Optics and Lasers in Engineering 135 (2020) 106187

Contents lists available at ScienceDirect

Optics and Lasers in Engineering

journal homepage: www.elsevier.com/locate/optlaseng

Transport of intensity equation: a tutorial

Chao Zuo a,b,∗, Jiaji Li a,b, Jiasong Sun a,b, Yao Fan a,b, Jialin Zhang a,b, Linpeng Lu a,b,
Runnan Zhang a,b, Bowen Wang a,b, Lei Huang c, Qian Chen b,∗
a
Smart Computational Imaging (SCI) Laboratory, Nanjing University of Science and Technology, Nanjing, Jiangsu Province 210094, China
b
Jiangsu Key Laboratory of Spectral Imaging & Intelligent Sense, Nanjing University of Science and Technology, Nanjing, Jiangsu Province 210094, China
c
Brookhaven National Laboratory, NSLS II 50 Rutherford Drive, Upton, NY 11973-5000, United States

a r t i c l e i n f o a b s t r a c t

Keywords: When it comes to “phase measurement” or “quantitative phase imaging”, many people will automatically connect
Transport of Intensity Equation (TIE) them with “laser” and “interferometry”. Indeed, conventional quantitative phase imaging and phase measurement
Quantitative Phase Imaging (QPI) techniques generally rely on the superposition of two beams with a high degree of coherence: complex interfero-
Phase Retrieval
metric configurations, stringent requirements on the environmental stabilities, and associated laser speckle noise
Partial Coherence
severely limit their applications in optical imaging and microscopy. On a different note, as one of the most well-
Optical Diffraction Tomography (ODT)
known phase retrieval approaches, the transport of intensity equation (TIE) provides a new non-interferometric
way to access quantitative phase information through intensity only measurement. Despite the insufficiency
for interferometry, TIE is applicable under partially coherent illuminations (like the Köhler’s illumination in a
conventional microscope), permitting optimum spatial resolution, higher signal-to-noise ratio, and better image
quality. In this tutorial, we give an overview of the basic principle, research fields, and representative applica-
tions of TIE, focus particularly on optical imaging, metrology, and microscopy. The purpose of this tutorial is
twofold. It should serve as a self-contained introduction to TIE for readers with little or no knowledge of TIE. On
the other hand, it attempts to give an overview of recent developments in this field. These results highlight a new
era in which strict coherence and interferometry are no longer prerequisites for quantitative phase imaging and
diffraction tomography, paving the way toward new generation label-free three-dimensional microscopy, with
applications in all branches of biomedicine.

1. Introduction
In physics and mathematics, the “phase” of a periodic signal is a
real-valued scalar that describes the relative location of each point on
the waveform within the span of each full period. The phase is typi-
cally expressed as an angle (in degrees or radians), in such a scale that
it varies by one full turn as the variable goes through each period. In
the branch of wave optics, the concept of phase is only limited to the
monochromatic coherent optical field, which can be described in terms
of a two-dimensional (2D) complex exponent with its argument so-called
the “phase” [1,2]. In comparison to phase, the amplitude component of
an optical field is generally much easier to understand or comprehend.
Because the square of the amplitude, also known as the “intensity” of
light (is also referred to as “irradiance” in radiometry [3]), is the only
visible component we can see with our eyes, representing the energy of
the light [4]. Human eyes or the imaging sensors can only detect the
intensity (or amplitude) component of light but the phase information
cannot be accessed directly. One important reason is that the oscillation
frequencies of light waves are around 1014 Hz [5], which is much higher
than the response speed of human eye (usually 30Hz) or photoelectric
imaging device (the frame rate of the current high-speed camera can
only reach 108 Hz level [6]).
The importance of phase can be illustrated by a simple and interest-
ing example in Oppenheim’s article [7] (see Fig. 1): Given two differ-
ent images, if we exchange their phase components in the frequency
domain while keeping their amplitude components unchanged, then
Fourier transforming their spectra back to spatial domain, the general
appearance of the two images will almost interchange. This simple ex-
ample shows that it is not the amplitude but the phase that bears most
of the information in the image (note that this example may not be very
appropriate because the phase here is the phase component of the im-
age spectrum rather than that of the optical field). The importance of
phase is particularly prominent in some specific fields, such as optical
metrology, material physics, adaptive optics, X-ray diffraction optics,
electron microscopy, and biomedical imaging. Most samples of interest
belong to phase objects with very little absorption, but the spatial dis-
tribution of their refractive index (RI) or thickness is nonuniform. There-
fore, they have small amplitude variations but strong phase components.

∗
Corresponding authors.
E-mail addresses: zuochao@njust.edu.cn (C. Zuo), chenqian@njust.edu.cn (Q. Chen).

https://doi.org/10.1016/j.optlaseng.2020.106187
Received 23 December 2019; Received in revised form 7 April 2020; Accepted 11 May 2020
Available online 19 June 2020
0143-8166/© 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
C. Zuo, J. Li and J. Sun et al.
Fig. 1. A example illustrating the importance of phase of an image. A exchange
of phase components in the frequency domain leads to the interchange of their
general appearance.
Fig. 2. The detector and the human eye are only sensitive to inten-
sity/amplitude while phase information is lost.
Human eyes or photodetectors are only sensitive to the amplitude vari-
ation rather than the phase variation, so they cannot “see” phase object
directly (Fig. 2). In other words, it is impossible to distinguish different
components of the object with different thickness or RI. Thus, the ac-
quisition of phase information is of particular importance in such cases.
Phase retrieval is a key research subject in the fields of optical
metrology and imaging technologies. It plays an indispensable role in
both industrial inspection and biomedical imaging applications. The
phase imaging techniques, especially for biological samples and weak-
absorptive semi-transparent objects, have a long history of develop-
ment. The cytoplasm and most organelles of biological cells have very
weak absorption, so they produce very little contrast under normal il-
lumination in a traditional bright-field microscope. To overcome this
difficulty, the most common approach is staining or labeling. Differ-
ent components of the cell have various affinities to different dyes or
fluorophores, producing enough intensity contrast or emitting differ-
ent fluorescence spectra (Fig. 3). The most common modality utilizing
exogenous contrast is fluorescence microscopy [8], in which a speci-
men is labeled with specific fluorescent molecules to provide targeted
morphological information. The structure of the organelle and protein
complex can also be visualized in three dimension (3D) with the help
of the optical sectioning capability of laser confocal microscopy and
multi-photon microscopy [9,10]. With the emergence of new fluorescent
molecular probes and the improvement of optical imaging technologies,
researchers have developed a variety of super-resolved fluorescence
microscopic methods that exceed the resolution limit of conventional
confocal microscopy, e.g., structured illumination microscopy (SIM)
[11], stimulated emission depletion microscopy (STED) [12], photo-
activated localization microscopy (PALM) [13], and stochastic optical
Optics and Lasers in Engineering 135 (2020) 106187
Fig. 3. Two typical cell labeling techniques. (a) chemical staining turns trans-
parent phase objects into absorptive objects; (b) fluorescent labeling turns trans-
parent phase objects into self-luminous objects.
reconstruction microscopy (STORM) [14]. These methods utilize the ex-
tremely high localization accuracy of single-molecule imaging or the
fluorescence excitation and bleaching properties of photoactivatable
fluorescent protein, significantly breaking the resolution limit of con-
ventional optical microscopy and revealing nanoscale structures of flu-
orescently labeled live cells. The relative importance of fluorescence
microscopy has recently been made evident by the 2014 Nobel Prize
in chemistry for “the development of super-resolved fluorescence mi-
croscopy”. However, such methods still require fluorescent dyes and
fluorescent proteins as biomarkers and are thus ill-suited for samples
that are non-fluorescent or cannot be easily fluorescently tagged. Be-
sides, the photobleaching and phototoxicity of the fluorescent agents
prevent live cells imaging over extended periods of time [15].
Although most biological cells are transparent and do not change
the amplitude of the light passing through them, they introduce phase
delays due to different structural regions of nonuniform optical density
(RI). In 1942, Zernike [16] invented the Zernike phase contrast (ZPC)
microscopy [Fig. 4(a)] to visualize phase optically. The method uses a
phase mask to shift only the unperturbed incident field component by
a quarter of a wavelength, such that it interferes with scattered field
with higher spatial frequencies, rendering the maximum contrast on the
interference image. This provides a simple, efficient method for con-
verting phase difference into intensity contrast, thus greatly improves
the contrast of the transparent phase object under an optical micro-
scope [Fig. 4(b)]. This led to a revolution in biological imaging, as
the significant contrast gain enabled the observation of unstained bi-
ological cells and tissues that were nearly invisible before, and the in-
vention earned Zernike the 1953 Nobel Prize in physics. Subsequently,
Nomarski [17] invented the Differential Interference Contrast (DIC) mi-
croscopy based on the principle of polarization beam-splitting and shear
interferometry in 1952. The phase contrast in DIC microscopy is propor-
tional to the phase gradient of the sample along the shear direction, giv-
ing the appearance of a 3D physical relief corresponding to the variation
of the optical density of a sample [Fig. 4(d)].
Although ZPC and DIC represent significant advancements in mi-
croscopy, these methods all suffer from one important drawback, that
the measured intensity has a nonlinear, and thus non-invertible, rela-
tionship with the phase of the specimen. Without the phase information,
morphologically relevant quantities of the sample, such as size, optical
thickness, dry mass density, and RI cannot be quantified. On the other
hand, “halo” [Fig. 4(b)] and “shadow off” [Fig. 4(d)] artifacts accom-
panied by these phase contrast methods also complicate the subsequent
image analysis and processing (such as cell counting and segmentation).
These limitations, when combined with the advent of digital image sen-
sors and advances in information optics, have resulted in the emerging
and burgeoning field of phase measurement and quantitative phase imag-
ing (QPI) [18,19], which combines innovations in optics, imaging the-
ory, and computational methods to image phase information of the sam-
ple quantitatively. Quantitative phase profiles of cells allow the deter-
mination of cellular structure and biophysical parameters with minimal
sample manipulation. Especially in cases where conventional prepara-
tion techniques, such as fixation, staining, or fluorescent tagging, may
C. Zuo, J. Li and J. Sun et al.
Fig. 4. Zernike phase contrast microscopy and differential interference contrast
microscopy and their imaging results on unstained check cells. (a) Zernike phase
contrast microscopy; (b) Zernike phase contrast microscopic image of unstained
check cells; (c) DIC microscopy; (d) DIC microscopic image of unstained check
cells.
affect cellular functions and limit biological insight, QPI offers an im-
portant alternative.
With the invention of the laser and the development of optical
metrology for nearly half a century, the most well-established phase
measurement method is through interferometry. Since the principle of
optical coherent interference was first proven to be a potential mea-
surement tool in the 1880s, interferometry has played a prominent role
in optical metrology. Nowadays, a high-precision laser interferometer
allows extremely sensitive phase measurements with the optical path
measurement accuracy up to 𝜆/100. The famous Laser Interferometer
Gravitational-wave Observatory (LIGO) [20,21] that detected gravita-
tional waves in 2016 is essentially a giant Michelson interferometer
with an arm span of 4 km (Fig. 5), enabling it to detect a minute distance
Fig. 6. Digital holography microscopy. (a) Schematic diagram of a typical digital holography microscope with a Mach-Zehnder configuration; (b) The basic principle
of the digital recording and numerical reconstruction.
Optics and Lasers in Engineering 135 (2020) 106187
Fig. 5. Laser Interferometer Gravitational-wave Observatory (LIGO) in
Louisiana and Washington, US.
change in length on the order of 1/10,000th the width of a single proton,
representing the incredibly small scale of the effects imparted by pass-
ing gravitational waves. Despite the advances of optical interferometry
over the decades, its basic principle remains unchanged: by superim-
posing an additional coherent reference beam with the original object
beam, the invisible phase information can be converted into a kind of
intensity signal, i.e., the interference pattern, which can be directly cap-
tured by traditional imaging devices. Through different types of fringe
analysis algorithms, the phase can be demodulated from the interfero-
gram. After decades of development, the interferometry has grown more
matured and derived many branches, such as electronic speckle pattern
interferometry (ESPI) [22,23], interferometric microscopy [24–26], and
digital holography [27–31]. Their basic principles are very similar, and
their developments are almost parallel. In particular, digital holography,
due to its unique advantages and flexibilities of digital recording and nu-
merical reconstruction, has made remarkable progress in the past two
decades and has become a benchmark for quantitative phase measure-
ment and microscopy [32–35] (Fig. 6). Nevertheless, the interferometric
C. Zuo, J. Li and J. Sun et al.
Fig. 7. Two typical wavefront sensing techniques. (a) Shack-Hartmann wavefront sensor; (b) pyramid wavefront sensor.
quantitative microscopic imaging represented by digital holography has
not shaken the position of traditional microscopic imaging technique,
failing to trigger an expected technological revolution in life science.
The main reasons lie in the interferometric characteristics of “quantita-
tive phase measurement”:
1) Digital holographic microscopy generally relies heavily on the light
sources with a high degree of temporal coherence (e.g., laser) and
spatial coherence (e.g., pinhole filtering) as well as complex inter-
ferometric optical configurations (i.e., object beam and reference
beam);
2) Due to the laser illumination sources typically used, digital holo-
graphic microscopy suffers from speckle noise arising from stray in-
terferences from imperfections in the optical system, which not only
limits the imaging resolution but also strongly deteriorates the image
quality;
3) Because of the high degree of spatial coherence of the light source,
the imaging resolution is only limited to the coherent diffraction
limit (half of the incoherent diffraction limit as in traditional inco-
herent microscope);
4) The additional reference beam path makes the measurement ex-
tremely sensitive to external disturbance (e.g., vibration, rapid air-
flow, etc.);
5) The phase demodulated from the interferogram is wrapped in the
range of (-𝜋, 𝜋] (so-called “wrapped phase”), and additional phase
unwrapping is needed to obtain the true absolute phase distribution.
In order to address or partially alleviate the above limitations of tra-
ditional interferometric phase imaging techniques, in recent years, many
researchers have gradually turned their attention to low-coherence
holography or white-light interferometric microscopy, e.g., spatial light
interference microscopy (SLIM) [36], white-light diffraction phase mi-
croscopy (wDPM) [37], quadriwave lateral shearing interferometry
(QWLSI) [38], and 𝜏 interferometry [39]. The combination of broadband
illumination with the common-path geometries significantly alleviates
the coherent noise problem and enhances the stability to mechanical
vibrations and air fluctuations that typically affect any interferometric
systems. However, most of them require relatively complicated optical
systems which are not typically available to most bio/pathologists, pro-
hibiting their widespread use in biological and medical science.
On a different note, the other broad category of phase measure-
ment techniques do not rely on interferometry, which is known as non-
interferometric phase measurement. A major branch of interferomet-
ric phase measurement techniques is called wavefront sensing, such as
Shack-Hartmann wavefront sensor [40–42] [Fig. 7(a)], pyramid wave-
front sensor [43–45] [Fig. 7(b)], and model-based wavefront sensor
[46,47]. The Shack-Hartmann wavefront sensor is a geometric optics
based phase measurement technique often used in adaptive optics sys-
tems with the main applications in telescope and astronomical observa-
tion. In such areas, “phase” is often referred to as the “wavefront aber-
Optics and Lasers in Engineering 135 (2020) 106187
ration”, representing the difference between the reference wavefront of
an ideal optical system and the distorted wavefront of an actual opti-
cal system. The original prototype of the Shack-Hartmann wavefront
sensor is an array of holes in a mask (so-called Hartmann screen) that
was developed by the astronomer Hartmann [41] in 1900. Although
Hartmann’s method has a simple design, its measurement accuracy and
light utilization efficiency are quite low. Later, in 1971, Shack and Platt
[40] modified the Hartmann screen by replacing the apertures in an
opaque screen with an array of lenslet, which focuses incoming ra-
diation to a spot array on the sensor [see Fig. 7(a)]. The integrated
gradient of the wavefront across the lenslet is proportional to the dis-
placement of the centroid relative to the reference regular spacing spot
array (ideal aberration-free case), and then the wavefront (phase) dis-
tribution can be reconstructed through integration. Nowadays, the ap-
plication of Shack-Hartmann wavefront sensor has expanded from as-
tronomical imaging to beam quality diagnosis [48,49], optical testing
[50,51], atmospheric laser communication [52,53], microscopic aber-
ration correction [54,55], ophthalmology [56–58] etc.. However, due to
the limitation of the physical size of the microlens, the Shack-Hartmann
sensor does not make full use of all the pixel resolution of the image
sensor, resulting in low spatial resolution of the recovered wavefront.
Consequently, the Shack-Hartmann wavefront sensor is seldom directly
used as an imaging tool, especially in the fields of QPI and optical
microscopy.
Another very famous and widely used non-interferometric phase
measurement technology is called phase retrieval. Since it is difficult
to measure the phase of the light directly but very easy to measure the
amplitude/intensity, we can consider the process of recovering (esti-
mating) the phase from the intensity distribution as a mathematical “in-
verse problem”, and the solution to such kind of problem is so-called
phase retrieval. Phase retrieval techniques can be divided into two cate-
gories: iterative methods and deterministic methods. The iterative phase
retrieval has its origins in ideas in electron microscopy. In 1972, Ger-
chberg and Saxton [59,60] proposed the first iterative phase retrieval
algorithm called Gerchberg-Saxton (GS) algorithm, which employs an it-
erative method in which the phase distribution within an object is found
that is consistent with both an intensity image of it and with its far-field
diffraction pattern. The proposed algorithm has been found to be able to
converge on the correct solution, and subsequently, the underlying ideas
have been applied to a wide variety of fields such as X-ray diffraction
imaging [61–64] (also referred to as coherent diffraction imaging, CDI)
[Fig. 8(a)], adaptive optics [65,66], and quantitative phase microscopy
[67–71]. A practical issue in X-ray imaging is that the object scattering
is so weak that the unscattered beam will swamp the signal of interest.
This problem is mitigated by the introduction of a beam stop to prevent
the undiffracted beam from destroying the detector at the expense of a
substantial low spatial frequency information loss. Miao et al. [61] cir-
cumvented this problem by using an additional electron microscope im-
age of the object to recreate the missing low spatial frequency data, and
C. Zuo, J. Li and J. Sun et al.
Fig. 8. Schematic layout of two main iterative phase retrieval methods. (a) Co-
herent diffraction imaging: a plane wave illuminates a sample, and the far-field
diffraction pattern is measured by a detector; (b) Ptychography: a coherent
probe is generated by an aperture of focusing optics. An extended sample is
scanned through the probe on a 2D grid, and diffraction patterns are collected
from a series of partially overlapping regions.
experimentally demonstrated the validity of this approach with X-ray
data. The rapid growth of interest in CDI has largely been driven by the
work of Miao et al. [61].
However, the GS algorithm also has some problems and limitations.
Some of the problems may result from the iterative algorithm itself, for
example, the convergence of the algorithm stagnates after early itera-
tions or even get trapped into local optima [72–74]. Various modified
algorithms have been proposed to improve the convergence of the orig-
inal GS algorithm. In 1973, Misell [75] pointed out that the GS algo-
rithm is not only applicable for the intensity distributions of the image
plane and the far-field Fourier plane but also can be extended to two
or more diffraction patterns of different propagation distances. He then
used more intensity constraints to improve the accuracy and conver-
gence of the iterative algorithm. Compared with the original GS algo-
rithm, Misell’s modified algorithm is more flexible and practical, which
opens up new ideas for improving the iterative phase retrieval algo-
rithm. Up to now, new improved and optimized algorithms have still
been emerging with a variety of techniques being used, such as mul-
tiple propagation distances [76,77], multiple illumination wavelengths
[78], phase modulation [79,80], and speckle illumination modulation
[81,82]. Other problems of the GS algorithm are inherent in solving in-
verse problems, such as the existence and uniqueness of the solution
[83–86]. Clearly, if an object is so small that a near-field image cannot
be acquired (only the far-field diffraction pattern is available), then re-
construction via the GS algorithm is not possible. In general, there is
no unique mapping between a 2D signal and its Fourier magnitude (dif-
ferent objects could produce identical far-field diffraction patterns) and
therefore, the problem is ill-posed. Bates [83] considered this problem
and claimed (without proof) that an object can be uniquely determined
by its autocorrelation function (i.e., the far-field diffraction pattern) if
its spatial extent is known (expect for several unimportant cases, such as
translation, phase conjugation, and inversion). In 1982, after carefully
analyzing the optimization principle behind the GS algorithm, Fienup
[87] pointed out that GS is equivalent to an error reduction algorithm,
which is essentially the same as the steepest descent algorithm. Fienup’s
work [87] has laid a solid theoretical foundation for the solvability of
the phase retrieval problem from the perspective of convex optimiza-
tion. It also provides a reasonable explanation for the stagnation prob-
lem: due to the non-convexity of the phase retrieval problem itself, there
Optics and Lasers in Engineering 135 (2020) 106187
is no guarantee that the error reduction algorithm can finally reach the
global minimum. Fienup [87,88] demonstrated that a modification of
the GS algorithm based on the theories of cybernetics, so-called hybrid
input-output (HIO) algorithm, could help prevent the iterative algorithm
from getting stuck in local minima where only support information is
used at the object plane. This algorithm has led to a number of analyses
and improvements, primarily in the manner in which the object esti-
mate is updated. These include the hybrid projection reflection [89],
difference map [90], relaxed averaged alternating reflections [91], and
charge flipping [92]. Though all of these ideas seem simple to imple-
ment, their success in application to experimental data requires con-
siderable skill and experience. A unified analysis of the methods has
been published by Marchesini [93], suggesting that none of them have
yielded an algorithm that has been found to converge in all cases. An-
other important innovation proposed by Marchesini et al. [63] is the
so-called “shrink-wrap” algorithm, in which the support for the object
is refined dynamically. This method is very useful for reconstructing the
atomic distribution within nanocrystals.
Ptychography is another iterative phase retrieval approach that has
its roots in electron microscopy [94]. In 2004, Rodenburg and Faulkner
[95,96] introduced the idea of “ptychography” into the GS and HIO al-
gorithms and presented a new iterative phase retrieval method so-called
Ptycholographic Iterative Engine (PIE). In the implementation of ptychog-
raphy, the object is illuminated by a field that is limited in extent and
the diffraction pattern is recorded, as illustrated in Fig. 8(b). This pro-
cess is repeated until the area of interest has been completely scanned.
The resultant spatially overlapped “sub-aperture” diffraction patterns
are used as intensity constraints, and the complex amplitude distribu-
tion of the area covered by the beam can be reconstructed by iterative
phase retrieval algorithm that is similar to GS and HIO. Compared with
GS and HIO, the ptychographic phase retrieval is more stable and re-
liable due to the high data redundancy resulting from the large num-
ber of spatially overlapped intensity data collected. Consequently, the
non-convexity of the phase retrieval problem is effectively alleviated,
and the ambiguity between the true solution and its conjugate is elim-
inated [86]. In the ensuing decade, the ptychographic phase retrieval
methods have been extensively studied and improved, and significant
progress has been made in probe (illumination function) recovery [97–
102], positional error correction [103–105], coherent mode decomposi-
tion [106–109], resolution improvement [110–113], and depth section-
ing for multi-layer samples [114–117]. It has now become an important
approach for high-resolution X-ray diffraction imaging [118–120], elec-
tron microscopy [121–123], and optical phase imaging and microscopy
[124–126].
In 2013, Zheng et al. [127] extended ptychography from the spa-
tial domain to frequency domain, presenting its dual counterpart named
Fourier ptychographic microscopy (FPM). In FPM, “ptychography” exists
in the Fourier domain: the sample is no longer scanned in the spatial
domain by an aperture-limited illumination beam, but is successively
illuminated by plane waves from different angles (usually with an LED
array); the intensity images are not acquired in the far field (Fourier
domain) but are low-resolution in-focus images captured directly in the
spatial domain. The finite aperture of the imaging system becomes the
“sub-aperture” in the frequency domain, while the changes of illumi-
nation angle enable the overlapping scanning of the sub-aperture in the
frequency domain, as illustrated in Fig. 9. Because of the perfect duality,
the key issues in FPM, such as aberration correction [128], positional
error correction [129,130], coherent mode decomposition [131–133],
depth sectioning for multi-layer samples [134,135], have been succes-
sively addressed in just a few years. At present, it has been developed
almost in parallel with conventional (spatial) ptychography. The ele-
gance of FPM lies in the fact that it not only retrieves the phase infor-
mation of the sample but achieves synthetic aperture naturally along
with the ptychographic updating process in the frequency domain, im-
proving the imaging resolution effectively [130,136,137]. In general, a
low-magnification objective with a low numerical aperture (NA) is used
C. Zuo, J. Li and J. Sun et al.
Fig. 9. Basic principle of Fourier ptychographic microscopy. The sample is suc-
cessively illuminated from different directions with different LED elements in
an LED array; the captured low-resolution images allow for simultaneous phase
retrieval and synthetic aperture in the frequency domain.
in FPM to obtain an inherent large field of view (FOV), and then the
high-angle illuminations are used to synthetic aperture. The final effec-
tive NA is the sum of the NA of the objective lens and the NA of the
illumination at the largest angle, and thereby, FPM generally achieves
a very large imaging space-bandwidth product (SBP). For example, in
2017, Sun et al. [138] reported a high-NA illumination based resolution-
enhanced FPM (reFPM) platform, in which a LED-array-based digital oil-
immersion condenser is used to create high-angle programmable plane-
wave illuminations, endowing a 10 × , 0.4 NA objective lens with the
effective imaging performance of 1.6 NA (308 nm lateral resolution for
coherent imaging) across a wide FOV of 2.34 mm2 (Fig. 9).
As mentioned above, the iterative phase retrieval effectively over-
comes the problem of “direct phase reconstruction from intensity mea-
surements”, and promotes the evolution of phase measurement from
interferometric to non-interferometric. In particular, the defocus-based
iterative phase retrieval and ptychographic imaging have contributed
significantly to the development and progress of adaptive optics, X-ray
diffraction imaging, and electron microscopy. However, there are still
two limitations: (1) a large amount of intensity data is often required to
ensure stable convergence; (2) a large number of iterations are needed
to obtain a reliable solution. These two shortcomings preclude their use
for high-speed and real-time imaging applications. Besides, the phase
retrieval algorithms relate the scattered field at the object to the scat-
tered field at the detector via a Fourier transform, a relationship that is
only valid for an optical field that is completely coherent. Thus, a re-
liable reconstruction was critically dependent on a very high degree of
coherence [139], which is an assumption that does not hold for many
partially coherent sources.
In contrast to iterative methods, another category of phase retrieval
method uses propagation to recover phase directly, in a non-iterative
manner. This idea was arguably first proposed by Teague [140,141]. In
1982, he first pointed out that under the paraxial approximation, one
could employ the conservation of energy on propagation (Helmholtz
equation) to derive a second-order elliptic partial differential equation
that outlines the quantitative relationship between the variation of in-
tensity along the optical axis to the phase of the optical field at the
plane perpendicular to the optical axis [140]. Teague believed that the
equation could be used as a phase retrieval approach. He termed this
equation transport of intensity equation (TIE), and this will be the main
theme of this tutorial. In his latter paper in 1983, Teague [141] proposed
a tentative approach to the solution of the equation based on Green’s
Optics and Lasers in Engineering 135 (2020) 106187
Fig. 10. The representive applications of TIE in the fields of (a) adaptive optics
[148], (b) X-ray diffraction imaging [161], (c) neutron radiography [164], and
(d) electron microscopy [166].
functions. Different from the iterative phase retrieval methods, TIE does
not employ traditional scalar diffraction formula to recover the phase
iteratively. When the intensity distribution at the in-focus plane (mea-
sured directly) and the axial intensity derivative (estimated by the finite
difference between two defocused intensities) are known, the phase in-
formation can be retrieved directly (deterministically) by solving TIE
without iterative operations. Compared with traditional interferometric
methods, TIE has many unique advantages of being non-interferometric
(without using reference beam), simple calculation (without iterations),
capabilities to temporally/spatially partially coherent beams (e.g., LED
illumination, halogen lamp, and built-in Köhler illumination of conven-
tional microscopes), does not need phase unwrapping (directly obtain
the absolute phase), complicated optical setups, and stable measurement
environment. Shortly thereafter, in 1984, Streibl [142] found that TIE
is applicable to optical microscopy as a tool for phase-contrast imaging
of transparent phase objects. Since the numerical solution to TIE has
not been proposed at that time, he only demonstrated the phase gradi-
ent enhancement effect of unstained cells by taking the axial intensity
derivative. Nevertheless, this pioneering work laid a preliminary theo-
retical foundation for the subsequent extensive applications of TIE in
the field of optical microscopy. In 1988, Ichikawa and Takeda [143] ex-
perimentally verified TIE for the first time. They solved TIE by Fourier
transform method and obtained the quantitative phase distribution of
the one-dimensional sample. However, rather than solving the partial
differential equation, this method was only found to be a special case
of grating shear interferometry [144]. In parallel, the research team of
the University of Hawaii led by Roddier [144–148] was also exploring
the possibilities of using TIE to correct the wavefront distortion intro-
duced by atmospheric turbulence in adaptive optics [Fig. 10(a)]. Assum-
ing that the amplitude of the wavefield is almost uniform, TIE can be
simplified from a more complicated elliptic partial differential equation
into a standard Poisson equation. This Poisson equation reveals that the
axial intensity derivative of the optical field is proportional to its wave-
front curvature (second derivative, i.e., Laplacian), so this technique is
often called curvature sensing (CS). There is no doubt that Roddier’s re-
search work has greatly promoted the TIE phase retrieval method. His
contribution is not only applying TIE to adaptive optics but more impor-
tantly, he first proposed the numerical solutions to the simplified form of
C. Zuo, J. Li and J. Sun et al.
TIE based on successive over-relaxation (SOR) method [145] and iterative
Fourier transform method [149]. In 1988, Roddier estimated the axial
intensity derivative by measuring the in-focus and slightly defocused
intensity signals at the entrance pupil of a telescope and fed it directly
back to the circuit of the deformable mirror to correct the atmospheric
wavefront distortion in real-time [145,146]. Then, in 1991, Roddier
[150] constructed the 13-unit low-order curvature sensing adaptive op-
tics system and presented preliminary laboratory experimental results.
The defocusing was realized through the rapid vibration of a curvature
variable membrane mirror so that the overfocus and underfocus inten-
sities can be measured simultaneously. The wavefront curvature signal
was measured by an avalanche photodiode array, and the deformable
mirror was driven based on the response matrix control method. Subse-
quently, the curvature sensing has been widely adopted in the US [the
3.6m-aperture CFHT (Canada-France-Hawaii Telescope) [151] and the
8m-diameter Gemini North telescope [152]], Europe [the 8.2m-aperture
VLTI (Very Large Telescope Interferometer)] [153], Japan (the 8.2m-
aperture Subaru Telescope) [154], and other countries’ large-scale as-
tronomical telescopes.
In the mid-to-late 1990s, the theoretical aspect of TIE developed
very rapidly and made fruitful achievements, wherein the X-ray diffrac-
tion imaging research group of the University of Melbourne was the
primary driving force. In 1995, Gureyev et al. [155] first proven the
well-posedness and uniqueness of the solution to TIE based on strict
mathematics: if the intensity of the optical field is strictly greater than 0
(precluding the case of optical vortex) and under the appropriate bound-
ary conditions, the solution of TIE exists and is unique (for the Neumann
boundary condition, the solution is unique up to a trivial additive con-
stant). This work has laid the theoretical foundation for the subsequent
TIE phase retrieval and QPI applications. In 1996, Gureyev and Nugent
[156,157] proposed to solve TIE based on the orthogonal expansion,
such as Zernike polynomial expansion or Fourier expansion, and pointed
out that when the intensity distribution is uniform, TIE can be solved
by using fast Fourier transform (FFT) efficiently [158]. Shortly thereafter,
Paganin and Nugent [159] extended the FFT-based solution so that it can
be effectively accommodated to the case of the nonuniform intensity dis-
tribution. This simple and effective FFT-based method has now become
the most widely used numerical solution to TIE. The establishment of
the mathematical basis and the numerical solution of TIE have become
a powerful vehicle that motivates its rapid development and widespread
applications in different fields. In 1995, Snigirev et al. [160] recognize
that propagation-based non-iterative phase retrieval techniques could
be applied in the field of X-ray imaging. Soon after, in 1996, Nugent
[161] et al. reported the quantitative phase image of a thin carbon foil
using 16 keV hard X-rays based on TIE [Fig. 10(b)]. These papers trig-
gered the rapid growth of interest of TIE in the applications of X-ray
optics [162,163]. Subsequently, the applications of TIE have also been
extended to the fields of neutron radiography [164,165] [Fig. 10(c)]
and transmission electron microscopy (TEM) [166–174] [Fig. 10(d)]. The
common feature of these fields is that it is difficult to obtain a highly co-
herent light source so that TIE became a simple and effective alternative
to interferometry.
In the late 1990s, on the one hand, the theoretical framework of TIE
was well established; on the other hand, researchers engaged in the field
of QPI were increasingly aware of the importance of using low coher-
ence illumination for improving the resolution and image quality. These
two aspects then became the catalyst for the widespread application of
TIE in the field of optical imaging and microscopy. In 1998, Paganin
and Nugent [159] reinterpreted the meaning of “phase” of partially co-
herent fields and pointed out that it is a scalar potential whose gradient
corresponds to the time-averaged Poynting vector. The importance of
this work is that it gives a more generalized and meaningful definition
of the “phase”, laying a theoretical basis for the subsequent TIE phase
retrieval under partially coherent illuminations. In the same year, Barty
et al. [175] briefly reported the QPI results of cheek cells and optical
fibers based on TIE, which was the first appearance of TIE in quantita-
Optics and Lasers in Engineering 135 (2020) 106187
tive optical phase-contrast imaging. Then Barty et al. [176] extended this
method to quantitative phase tomography, based on a similar approach
as the “holotomography” earlier demonstrated by Cloetens et al. [163] in
1999. Owing to its successful applications in the field of QPI and opti-
cal microscopy, at the beginning of the 20th century, the theory of the
TIE phase retrieval happened to coincide with the optical transfer func-
tion (OTF) theory of optical microscopy. In 2002, Barone-Nugent et al.
[177] theoretically analyzed the image formation of weak phase objects
under an optical microscope based on the 3D OTF theory developed by
Striebl [178], further confirmed the applicability of TIE under partially
spatially coherent illuminations. In the same year, Sheppard [179] an-
alyzed the TIE phase retrieval under the first-order Born approximation
and pointed out that the low-frequency component of the axial intensity
derivative is proportional to the phase Laplacian (second-order deriva-
tive). The conclusion is consistent with the case of the simplified version
of TIE under uniform intensity analyzed by Roddier [144–146]. The ini-
tial establishment of TIE in the field of optical microscopy and the ex-
perimental results of Barty et al. led the application of TIE into the fields
of biological and biomedical imaging. Curl et al. [180,181] used TIE to
quantitatively detect the morphology and growth rate of cells, indicating
that quantitative phase information is very conducive to data processing
and quantitative analysis, such as cell segmentation and counting. Ross
et al. [182] utilized TIE to enhance the imaging contrast of unstained
cells under microbeam irradiations. Dragomir et al. [183] successfully
applied TIE to the quantitative measurement of the birefringence effect
of unstained isolated cardiomyocytes. Not just limited to biological sam-
ples, TIE was also successfully applied to the quantitative measurement
of phase and RI distribution of optical fibers and other technical phase
objects [184–187].
From 2010 to date, the TIE phase retrieval reached another prime
due to the brilliant achievements in visible light imaging and mi-
croscopy. During this period, the basic theory of TIE (especially for the
theory of partially coherent imaging) and related applications (espe-
cially for the applications of quantitative phase microscopy and tomog-
raphy) have made a great leap forward again, reaching new and un-
precedented heights. Kou et al. [188] demonstrated that TIE could be
directly implemented based on an off-the-shelf DIC microscope to re-
alize QPI, and successfully obtained the quantitative phase images of
human cheek cells. In the same year, Waller et al. [189] proposed a
method named high-order TIE by using more intensity measurements
at multiple defocused planes to enhance the accuracy and robustness
of phase retrieval. The simple and straightforward descriptions of these
work greatly boosted the popularity of TIE in the fields of optical imag-
ing and microscopy. Meanwhile, Waller et al. also proposed single-shot
TIE optical configurations based on volume holography [190] and chro-
matic aberration [191], enabling acquisition of intensity images at dif-
ferent defocus distances from a single exposure. In 2011, Kou et at.
[192] put forward an OTF deconvolution method for QPI under par-
tially coherent illuminations and quantitatively compared it with the
conventional TIE. In 2012, Almoro et al. [193] used speckle field illumi-
nation to enhance the phase contrast of the defocused intensity images
for smooth objects. In the same year, Gorthi and Schonbrun [194] first
applied TIE to flow cytometry and realized the automatic collection of
through-focus intensity images by tilting the microfluidic channel verti-
cally. In the meantime, there are a large number of publications focusing
on optimizing the finite-difference scheme phase retrieval to improve
the accuracy and robustness of TIE under noisy conditions [195–200].
Since 2012, our group has carried out systematic research work on the
TIE phase retrieval, and several key theoretical issues in TIE have been
thoroughly and comprehensively studied:
1) Efficient numerical solution under inhomogeneous boundary condi-
tions [201–203];
2) Phase discrepancy analysis and compensation [204];
3) Optimal axial intensity derivative estimation [205,206];
4) Phase space extension for partially coherent fields [207,208];
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 11. Classification of the phase imaging techniques.

Fig. 12. Timeline of TIE research fields. Key developments in both theoretical and experimental aspects are shown.
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

5) Resolution enhancement via illumination engineering [209–211].

6) Diffraction tomography under partially coherent illuminations Table 1
[212–214]; List of acronyms used in the tutorial

Acronym Full Name

In experimental aspects, instead of including a mechanical refocus-
ing system (e.g., piezo-stage) to provide the through-focus scanning, we 1D One-Dimensional
2D Two-Dimensional
proposed to introduce an electrically tunable lens [215] or a spatial light
2.5D Two-Point-Five-Dimensional
modulator (SLM) [216] placed in the Fourier conjugated plane of the de- 3D Three-Dimensional
tector in a conventional wide-field transmission microscope. This pro- 4D Four-Dimensional
vides fast high-speed acquisition of through-focus intensity stack, mak- AOTF Amplitude Optical Transfer Function
ing the dynamic TIE phase imaging for live cells possible. We have AF Ambiguity Function
ATF Amplitude Transfer Function
successfully applied these systems for investigations of drug-induced BC Boundary Condition
morphology changes and phagocytosis of macrophages [216], imaging CARS Coherent Anti-Stokes Raman Scattering
of cellular dynamics of cancer cells [215,217], and characterization of CCF Complex Coherence Factor
micro-optical elements [202]. To date, the TIE phase retrieval attracts CDC Complex Degree of Coherence
CDI Coherent Diffraction Imaging
more and more attention from both academia and industry, and related
CT Computerized Tomography
new theories and technologies are still constantly emerging. CTF Contrast Transfer Function
As previously mentioned, we have briefly summarized the develop- CSD Cross-Spectral Density
ment of phase imaging and measurement technologies, suggesting that DC Direct Current
these technologies can be divided into two main categories: phase vi- DCT Discrete Cosine Transform
DIC Differential Interference Contrast
sualization method (phase-contrast imaging) and phase measurement DOE Diffractive Optical Element
method (quantitative phase imaging). The essential difference between DPC Differential Phase-Contrast
these two categories is whether they can recover the “quantitative” FFT Fast Fourier Transform
phase information. For the phase measurement, it can also be divided FOV Field of View
FPM Fourier Ptychographic Microscopy
into interferometry, wavefront sensing, and phase retrieval. Interfer-
FTS Fourier Transform Spectrometer
ometry and phase retrieval also have their independent sub-branches. GFST Geometric-Flow Speckle Tracking
Fig. 11 illustrates the classification of phase imaging and measurement GTIE Generalized Transport of Intensity Equation
technologies. In the following parts, the emphasis will be focused on GSM Gaussian-Schell Model
the non-interferometric phase measurement technology based on TIE. LHS Left Hand Side
LMI Light Field Moment Imaging
As discussed above, although the phase retrieval and quantitative phase MCF Mutual Coherence Function
microscopic imaging based on TIE have been extensively reported in MI Mutual Intensity
recent years, there has been no new comprehensive review of TIE focus- MSE Mean Square Error
ing on the basic principle, technical details, and practical applications NA Numerical Aperture
ODT Optical Diffraction Tomography
of this emerging technology. Therefore, at present, there appears to be a
OFS Optimal Frequency Selection
moment in time for this tutorial, which describes the fundamental prin- OTF Optical Transfer Function
ciples and latest progress of TIE in a rigorous and comprehensive way is PAM Partitioned/Programmable Aperture Microscopy
of interest to the community of optical scientists and engineers, physi- PGTF Phase Gradient Transfer Function
cists, and biologist interested in optical microscopy and phase imaging. PIE Ptycholographic Iterative Engine
POTF Phase Optical Transfer Function
Before we start, a timeline of TIE research developments is outlined
PTF Phase Transfer Function
in Fig. 12, which provides an overview of the relevant research fields PSD Power Spectral Density
and summarizes the key developments in both theoretical and exper- PSF Point Spread Function
imental aspects. We have also included Tables 1 and 2 at the begin- PST Phase Space Tomography
PWM Pulse Width Modulation
ning with a list of all the acronyms, notations, and symbols used in the
QPI Quantitative Phase Imaging
tutorial. RHS Right Hand Side
The remaining parts of this paper are organized as follows: In RI Refractive Index
Section 2, we start with the motivations behind the development of TIE, SBP Space-Bandwidth Product
describe its fundamental principle, outline its derivations, and physi- SCF Self-Coherence Function
SGDF Savitzky-Golay Differential Filter
cal implications. From Sections 3 to Section 8, several key theoretical
SNR Signal-to-Noise Ratio
issues of TIE are discussed in detail, including its solutions, image for- TCC Transmission Cross-Coefficient
mation model under coherent illumination, axial intensity derivative es- TEM Transmission Electron Microscopy
timation, QPI under partially coherent illuminations, and 3D diffraction TIDT Transport of Intensity Diffraction Tomography
TIE Transport of Intensity Equation
tomography based on the transport of intensity. In Section 9, the sys-
TV Total Variation
tem configurations and typical applications of TIE are presented with an WDF Wigner Distribution Function
emphasis on optical imaging, metrology, and microscopy in the visible WOTF Weak Object Transfer Function
light band. Finally, the challenging problems, as well as future research ZPC Zernike Phase Contrast
directions, are outlined in Section 10. GS Gerchberg-Saxton
HIO Hybrid Input-Output
PIE Ptycholographic Iterative Engine
2. Basic concept of TIE CS Curvature Sensing
SOR Successive Over-Relaxation
IFFT Inverse Fast Fourier Transform
TIE is a partial differential equation that describes the quantitative NRR Noise Reduction Ratio
relationship between the phase and the axial intensity variation of the GP Gaussian Process
optical field. In this section, we will start from the simplest physical ROC Radius of Curvature
phenomenon without invoking complicated mathematics. Then we will LUT Lookup Table
MFPI Multi-Filter Phase Imaging
use plain words to explain the transport of intensity phenomenon and
RMSE Root Mean Square Error
its cause (Subsection 2.1). In the following Subsection 2.2, we will use
three different approaches to derive TIE based on strict mathematical
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Table 2
List of notations and symbols used in the tutorial

Notation or Symbol Definition

𝐱 = (𝑥, 𝑦) Lateral 2D spatial coordinate (𝜇m)

z Longitudinal spatial coordinate (𝜇m)
𝐫 = (𝐱, 𝑧) 3D spatial coordinate (𝜇m)
𝐮 = (𝑢, 𝑣) Lateral 2D spatial frequency coordinate (𝜇𝑚−1 )
𝜂 Longitudinal spatial frequency coordinate (𝜇𝑚−1 )
𝐟 = (𝐮, 𝜂) 3D spatial frequency coordinate (𝜇𝑚−1 )
√
𝑟 = |𝐱| = 𝑥2 + 𝑦2 , 𝜃 = tan−1 (𝑦∕𝑥) 2D polar coordinates
√
𝜌 = |𝐮| = 𝑢2 + 𝑣2 , 𝜗 = tan−1 (𝑣∕𝑢) 2D spatial-frequency polar coordinates
k ≡ (kx , ky , kz ) 3D spatial frequency coordinate in k space
(𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ) ≡ (𝑢, 𝑣, 𝜂) 3D spatial frequency coordinate (𝜇𝑚−1 )
(𝑥, 𝑦, 𝑢, 𝑣) = (𝐱, 𝐮) 4D coordinate in phase space
𝜆 Wavelength (𝜇m)
k0 Wave number in vacuum (𝜇𝑚−1 )
km Wave number in medium (𝜇𝑚−1 )
n0 Background refractive index
n(r) Refractive index distribution of the object
Δn Refractive index difference between object and medium
j Imaginary unit
𝛿 Delta function
𝜔 Angular frequency
(𝛼, 𝛽, 𝛾) Direction angle
∇ Transverse gradient
Ω Open bounded domain in 2D space
𝜕Ω Piecewise smooth boundary of Ω
Ω̄ Closed bounded domain in 2D space
ℝ2 2D real space
ℝ𝑁 N-dimensional real space
∇x Gradient in 3D space
⊗ Convolution
⊗ 2D spatial convolution in phase space
𝐱
⊗ 2D spatial-frequency convolution in phase space
𝐮
⊗ 4D convolution in phase space
𝐱,𝐮
̂⋅ Fourier transform of the corresponding viable
̃⋅ Estimated value of the corresponding viable

and physical theories. At last, in Subsection 2.3, the mathematical ex-
pression of TIE and its physical meaning are explained to echo the earlier
Subsection 2.1.
2.1. Transport of intensity effect
Since the traditional photodetectors (e.g., CMOS and CCD) are only
sensitive to the intensity of the optical field and the phase information
cannot be directly detected, most of the phase measurement methods
need to convert the invisible phase information into the visible intensity
information for recording and reconstruction. For example, the classi-
cal interferometry transforms the invisible phase information into visi-
ble fringe patterns by spatial superposition of two coherent light beams,
and the fringe pattern can be then recorded by photodetectors. Finally,
the quantitative phase distribution can be demodulated from the inter-
ferogram through fringe analysis algorithms.
The phase-to-intensity conversion does not only exist in the interfer-
ence phenomena, and the propagation of the optical field is another kind
of spontaneous phase-to-intensity conversion process. We can imagine
that a swimming pool in the sunshine and the slight ripples appear on
the surface of the pool because of the breeze or a dropped swimming
ring. Looking towards the bottom of the pool, we can see the wave-like
pattern created by the refraction of the light at the pool surface (see left
of Fig. 13). The ripples on the water surface act as lenses in a variety
of shapes, and the transmitted light is redistributed and may be concen-
trated in some regions of the pool bottom, instead of evenly distributed
at the whole bottom. If the continuously changing curvature on the wa-
ter surface is large enough, the water wave will focus the sunlight onto
the bottom of the pool, forming some bright line network structure, as
shown in the middle figure of Fig. 13.
So, what is the relationship between the above phenomenon and
phase retrieval? In fact, the water surface with ripples is essentially a
phase object, and it only changes the phase of the incoming light. The
light-and-dark network structure at the bottom of the swimming pool is
exactly the “manifestation of phase via propagation”. We call this kind
of phenomenon “transport of intensity effect”. In this context, there is
neither laser nor interference, but the intensity pattern at the bottom of
the swimming pool closely resembles the interference pattern in optical
interferometry. This phenomenon suggests that the phase can be con-
verted into intensity without resorting to interferometry, and this kind
of intensity is often referred to as “phase contrast”. Similar to the sit-
uation that the light-and-dark intensity structure at the bottom of the
swimming pool reflects the shape of the water surface, the aim of the
non-interferometric TIE phase retrieval is to recover the quantitative
phase distribution by measuring the phase-induced intensity variation
(phase contrast signal) at the defocused plane. In this case, the mor-
phology of the water surface is retrieved from the intensity pattern at
the bottom of the swimming pool, as illustrated in the right image of
Fig. 13.
We are no strangers to the transport of intensity effect: the twin-
kling of stars in the night, the distorted scene outside the window in
the rain, and the above-mentioned network of bright lines at the bot-
tom of a swimming pool in the sunshine. These are all manifestations
of phase, implying the inextricable ties between phase and intensity of
the light wave. All these phenomena show that although phase objects
cannot be observed directly, they incessantly, and implicitly manifest
their existence. Nevertheless, the transport of intensity effect is not al-
ways observable. For instance, when the water surface is completely still
and almost flat, the intensity patterns will disappear. The transport of
intensity effect can be explained qualitatively based on geometric op-
tics, as shown in Fig. 14. If the observed optical field is a perfect plane
C. Zuo, J. Li and J. Sun et al.
Fig. 13. The wave-like pattern at the bottom of a swimming pool in sunlight. The pool surface refracts the incident sunlight to produce the light-and-dark network
structure at the bottom of the swimming pool.
Fig. 14. The influence of phase on the beam propagation. (a) For on-axis plane waves, the intensity does not change during propagation; (b) non-planar phase
change the intensity during propagation; (c) an intuitive geometric-optics interpretation of transport of intensity effect.
wave, the intensity will not change with the propagation distance Δz, as
shown in Fig. 14(a). This is because the plane wave can be regarded as
a cluster of parallel rays with the propagation direction parallel to the
optical axis (perpendicular to the wavefront, i.e., equiphase surface), as
shown in Fig. 14(c). Since light rays spread along straight lines in free
space, the intensity distribution maintains the same regardless of the lo-
cation. However, for non-planar waves, the intensity will change with
the propagation distance, as shown in Fig. 14(b). Fig. 14(c) gives the in-
tuitive geometric interpretation of this phenomenon. Though wavefront
distribution is not uniform, it can still be regarded as the combination of
several piece-wise uniform functions. In each small area, the distribution
of wavefront can be approximated as a plane wave, and the propagation
direction of light rays in these areas is perpendicular to their wavefront.
However, the phase distribution is not uniform for the whole region, and
the distribution and propagation direction of light rays in different ar-
eas may be different from each other. The convergence or divergence of
light rays in different areas will inevitably cause intensity variation dur-
ing the propagation. On the other hand, there is a certain requirement of
transport of intensity effect on the spatial coherence of the illumination.
If it is cloudy, and the sunlight can no longer illuminate the pool sur-
face directly due to the multiple scattering of the cloud, we can hardly
observe the transport of intensity phenomenon (the relevant part will
be discussed in detail in Section 6). Based on the transport of intensity
effect mentioned above, the quantitative relationship between the axial
intensity variation and the phase is established by TIE, and the quan-
titative phase distribution can be directly recovered by solving TIE. In
the next subsection, we will introduce TIE and its detailed derivation.
2.2. Derivation of TIE
In 1982, Teague [140,141] first established the quantitative rela-
tionship between the longitudinal intensity variation and phase of a
coherent beam with use of a second-order elliptic partial differential
equation, so called TIE. Considering a paraxial monochromatic coherent
beam propagating along z-axis, its complex amplitude can be expressed
√
as 𝑈 (𝐱, 𝑧) = 𝐼 (𝐱, 𝑧) exp [𝑗𝜙(𝐱)], and TIE can be expressed as
𝜕𝐼 (𝐱, 𝑧)
−𝑘 = ∇ ⋅ [𝐼 (𝐱, 𝑧)∇𝜙(𝐱)] (1)
𝜕𝑧
Optics and Lasers in Engineering 135 (2020) 106187
where k is the wave number 2𝜋/𝜆; x is the transverse coordinates
𝐱 = (𝑥, 𝑦); I(x, z) is intensity distribution at the plane located at the prop-
agation distance z, and ∇ is gradient operator (Hamiltonian) over x. In
Teague’s original TIE paper, he introduced the expression of complex
amplitude into Helmholtz equation under the paraxial approximation
(paraxial wave equation), and then separated the real part and the imag-
inary part to obtain TIE. In fact, TIE can be also derived from Poynting
theorem or Fresnel diffraction formula Under the paraxial approxima-
tion and the limit of small propagation distance. Next, we will show
three different ways to derive TIE.
2.2.1. Derivation of TIE from paraxial wave equation
TIE can be derived from the paraxial wave equation, which is also the
approach adopted by Teague [140,141] when TIE was first derived. We
know that the free space propagation of the monochromatic coherent
beam obeys Helmholtz equation [5],
( 2 )
∇ + 𝑘2 𝑈 (𝐫) = 0 (2)
where ∇ is the Hamiltonian in 3D space 𝐫 = (𝑥, 𝑦, 𝑧). The complex am-
plitude of coherent beams must satisfy Helmholtz equation, and plane
waves and spherical waves are the basic solutions to Helmholtz equa-
tion. Considering a paraxial monochromatic coherent beam propagating
along z-axis, it can be expressed as
𝑈 (𝐫) ≈ 𝑈 (𝐱, 𝑧)𝑒𝑗𝑘𝑧 (3)
The scalar complex amplitude of the paraxial optical field is denoted
as U(x, z), which is a slowly varying function with respect to z. Note
in Eq. (3), the time-dependent part 𝑒−𝑗𝜔𝑡 (has no effect on intensity)
has been omitted, where 𝜔 is the angular frequency of the light wave.
The paraxial wave equation can be obtained by substituting Eq. (3) into
Helmholtz equation
𝜕𝑈 (𝐱, 𝑧)
∇2 𝑈 (𝐱, 𝑧) + 2𝑗𝑘 =0 (4)
𝜕𝑧
It is the propagation law that the complex amplitude of a paraxial
monochromatic coherent beam must obey. Note ∇ in Eq. (4) is now
defined over x. Next,
√ we substituted the scalar complex amplitude ex-
pression 𝑈 (𝐱, 𝑧) = 𝐼 (𝐱) exp [𝑗𝜙(𝐱)] into the paraxial wave equation, and
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

separate the real part (in the following derivation, the spatial coordinate Eq. (12) suggests that the complex amplitude distribution in the diffrac-
is omitted for notation simplicity) tion plane is the weighted linear (interferential) superposition of spher-
( ) ical wavelets from all points r′ in the original plane. However, it is
( ) 𝜕𝑈 𝜕𝑈 ∗
∇ 𝑈 ∗ ∇𝑈 − 𝑈 ∇𝑈 ∗ + 2𝑗𝑘 𝑈 ∗ +𝑈 =0 (5) known that Kirchhoff’s boundary conditions lead to physical solutions
𝜕𝑧 𝜕𝑧
containing a grave mathematical inconsistency. Rayleigh and Sommer-
where feld [219,220] solved this problem by deriving two possible integration
∇𝑈 = 𝑗𝑈 ∇𝜙 (6) formulas with use of two different Green’s functions under“physically
correct” boundary conditions. The Rayleigh-Sommerfeld integrals for-
mulas only differ slightly with Kirchhoff’s formula, the resulting obliq-
𝑈 ∗ ∇𝑈 = 𝑈 ∗ 𝑗𝑈 ∇𝜙 = 𝑗𝐼∇𝜙 (7) uity factors become 𝐾 (𝜃) = cos 𝜃 for the first Rayleigh-Sommerfeld inte-
gral (RS-I type), and 𝐾 (𝜃) = 1 for the second Rayleigh-Sommerfeld in-
tegral (RS-II type). A comparison of these obliquity factors reveals that
𝑈 ∗ ∇𝑈 − 𝑈 ∇𝑈 ∗ = 2𝑗𝐼∇𝜙 (8)
the Kirchhoff’s solution is the arithmetic average of the two Rayleigh-
( ) Sommerfeld solutions [2,5]. The existence of obliquity factor suggests
( )
𝜕𝑈 𝜕𝑈 ∗ 𝜕 𝜕𝐼 that the amplitude contribution in the original plane of each spheri-
𝑗 𝑈∗ +𝑈 =𝑗 𝑈𝑈∗ = 𝑗 (9)
𝜕𝑧 𝜕𝑧 𝜕𝑧 𝜕𝑧 cal wavelet emitted from each point may not be isotropic. However, if
Substituting Eqs. (8) and (9) into Eq. (5), we can deduce TIE the diffraction angle is so small that the paraxial approximation is sat-
isfied, K(𝜃) ≈ 1 (regardless its form), the effect of the obliquity factor
𝜕𝐼
−𝑘 = ∇ ⋅ (𝐼∇𝜙) (10) can be neglected. Then the diffraction becomes a “linear shift invariant
𝜕𝑧
system”
It is a second order elliptic partial differential equation about the phase
function 𝜙. Similarly, “transport of phase equation” can be obtained 𝑈Δ𝑧 (𝐫) = 𝑈0 (𝐫) ∗ ℎ(𝐫) (14)
from the imaginary part of the paraxial wave equation The system is uniquely determined by its spatially invariant impulse
1 2 𝜕𝜙 1 response function in the spatial domain
−2𝑘𝐼 = 𝐼 ∇2 𝐼 − (∇𝐼 )2 − 𝐼 2 (∇𝜙)2 + 𝑘𝐼 2 (11)
𝜕𝑧 2 4 1 exp (𝑗 𝑘|𝐫 |)
ℎ (𝐫 ) = (15)
Because the phase is not directly accessible, this equation is difficult to 𝑗𝜆 |𝐫 |
be applied directly. We just give the expression of this equation here
Here we assume that the original plane and the diffraction plane are lo-
considering the completeness of this tutorial.
cated at (x, 0) and (x, z), respectively. The expression of the impulse re-
sponse function can be further simplified and expressed as the 2D scalar
2.2.2. Derivation of TIE from Fresnel diffraction formula form
The diffraction law of coherent light field can be accurately described ( √ )
in the spatial domain or in the spatial frequency domain from the view- exp 𝑗𝑘 Δ𝑧2 + |𝐱|2
point of “system”. In this subsection, we use the Fresnel diffraction for- 1
ℎΔ𝑧 (𝐱) = √ (16)
mula to derive TIE in the spatial and frequency domain, respectively. 𝑗𝜆
Δ𝑧2 + |𝐱|2
1) Using Fresnel diffraction integral to derive TIE in the spatial For a paraxial beam propagating along z-axis, the expression of the im-
domain pulse response function can be further simplified as follows
{ }
Based on the Huygens-Fresnel principle, Kirchhoff [218] derived the 1 𝑗𝜋
ℎΔ𝑧 (𝐱) = exp(𝑗 𝑘Δ𝑧) exp |𝐱|2 (17)
rigorous Kirchhoff diffraction formula based on the Helmholtz equation 𝑗𝜆Δ𝑧 𝜆Δ𝑧
and Green’s theorem, under certain boundary conditions at the diffrac-
Substituting Eq. (17) into Eq. (14), we can get the Fresnel diffraction
tion plane. The basic idea is based on the principle of linear superpo-
integral [2]
sition: since any complex light source can be regarded as a combina- { }
tion of individual point sources, the complex light waves can always be exp(𝑗𝑘Δ𝑧) 𝑗𝜋 |
𝐱 − 𝐱0 || 𝑑 𝐱0
2
𝑈Δ𝑧 (𝐱) = 𝑈0 (𝐱0 ) exp | (18)
decomposed into linear combinations of simple spherical waves. And 𝑗𝜆Δ𝑧 ∫ 𝜆Δ𝑧
the linearity of the wave equation allows each spherical wave to ap- If in addition to the Fresnel approximation the stronger Fraunhofer ap-
ply the above principles respectively. Then their contributions on the proximation is satisfied, then the quadratic phase factor inside the inte-
diffraction plane are superposed to create the resultant diffraction field. gral of Eq. (18) is approximately unity, and the diffraction field can
Considering that the complex amplitude of the light wave in 3D space be found (up to a multiplicative phase factor in x) directly form a
is 𝑈0 (𝐫) = 𝑎(𝐫)𝑒𝑗 𝜙(𝐫 ) (in order to simplify the representation, we denote Fourier transform of the original field. Thus in the region of Fraun-
𝐫 = (𝐱, 𝑧) as the 3D spatial coordinate vector), the complex amplitude hofer diffraction (or equivalently, in the far field), aside from multi-
of the optical field obtained after propagating a distance of Δz can be plicative phase factors preceding the integral, the diffraction field is sim-
expressed as ply the Fourier transform of the original field, evaluated at frequencies
( ) 𝐟 = 𝐱∕𝜆Δ𝑧.
𝑈Δ𝑧 (𝐫) = 𝑈0 (𝐫 ′ )ℎ 𝐫 ′ , 𝐫 𝑑 𝐫 ′ (12)
∫ Next, we derive TIE based on Fresnel diffraction integral. Consider-
ing a complex amplitude 𝑈0 (𝐱) = 𝑎(𝐱)𝑒𝑗𝜙(𝐱) of a paraxial monochromatic
The above formula shows that wave propagation can be regarded as a
coherent beam propagating along the optical axis, the distance between
linear system, and the impulse response function of the system h(r′, r)
the diffraction plane and the object plane is Δz. According to Eq. (18),
is exactly the complex amplitude on the diffraction plane generated by
the intensity distribution in the diffraction plane can be expressed as
the spherical wavelet propagated from r′
( ) [ )]
( ) exp 𝑗 𝑘||𝐫 − 𝐫 ′ || 1 𝑗𝜋 (|
𝐱 − 𝐱1 || − ||𝐱 − 𝐱2 ||
2 2
1 𝐼Δ𝑧 (𝐱) = 𝑈Δ𝑧 (𝐱)𝑈Δ𝑧
*
(𝐱 ) = 𝑈 (𝐱 )𝑈 * (𝐱 ) exp 𝑑 𝐱1 𝑑 𝐱2
ℎΔ𝑧 𝐫 ′ , 𝐫 = 𝐾 (𝜃) (13) 𝜆Δ𝑧 ∫ 0 1 0 2 𝜆Δ𝑧 |
𝑗𝜆 |𝐫 − 𝐫 ′ | [ ( )] ( )
1 𝑗𝜋 | |2 | |2 𝐱 − 𝐱1
= 𝑈 (𝐱 )𝑈 * (𝐱 ) exp 𝐱 − |𝐱 2 | exp 2𝜋𝑗𝐱 ⋅ 2 𝑑 𝐱1 𝑑 𝐱2
exp (𝑗 𝑘|𝐫 −𝐫 ′ |) 𝜆Δ𝑧 ∫ 0 1 0 2 𝜆Δ𝑧 | 1 | 𝜆Δ𝑧
where |𝐫 −𝐫 ′ |
is an ideal spherical wave emitted from r′ and prop-
(19)
agated to r. 𝐾 (𝜃) = 12 (1 + cos 𝜃) is the obliquity factor, where cos 𝜃 =
Δ𝑧
|𝐫 −𝐫 ′ |
is the angle between z-axis and the line connecting 𝐫 and 𝐫 ′ . Taking Fourier transform on Eq. (19), we get the intensity spectrum
in the spatial frequency domain [221]
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

{ } 1
𝐼̂Δ𝑧 (𝐮) = ℱ 𝐼Δ𝑧 (𝐱) = 𝑈 (𝐱 )𝑈 ∗ (𝐱 )
𝜆Δ𝑧 ∫ 0 1 0 2
[ ]
𝑗𝜋 (| |2 | |2 ) { ( 𝐱 − 𝐱1 )}
exp |𝐱1 | − |𝐱2 | ℱ exp 2𝜋𝑗𝐱 ⋅ 2 𝑑 𝐱1 𝑑 𝐱2
𝜆Δ𝑧 𝜆Δ𝑧
[ ( ) ] (
1 𝑗𝜋 | |2 | |2 𝐱 − 𝐱1 )
= 𝑈 (𝐱 )𝑈 ∗ (𝐱 ) exp 𝐱 − |𝐱2 | 𝛿 𝐮− 2 𝑑 𝐱1 𝑑 𝐱2
𝜆Δ𝑧 ∫ 0 1 0 2 𝜆Δ𝑧 | 1 | 𝜆Δ𝑧
𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮
𝐱1 =𝐱− 2 ,𝐱2 =𝐱+ 2 ( ) ( )
𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮
= 𝑈 𝐱− 𝑈0∗ 𝐱 + exp {−2𝑗𝜋𝐱 ⋅ 𝐮}𝑑𝐱
∫ 0 2 2
(20)

Note that the constant scale factor in above formula is omitted. When
the( propagation
) distance Δz → 0, the first-order Taylor expansion of
𝑈0 𝐱 − 𝜆Δ𝑧2
𝐮
can be written as
( )
𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮
𝑈0 𝐱 ± ≈ 𝑈0 (𝐱) ± ⋅ ∇ 𝑈 0 (𝐱 ) (21)
2 2
Substituting Eq. (21) into Eq. (20) and only retaining the linear term of
Δz [210,216]
Δ𝑧
𝐼Δ𝑧 (𝐱) = 𝐼0 (𝐱) − ∇ ⋅ [𝐼0 (𝐱)∇𝜙(𝐱)] (22)
𝑘
𝐼Δ𝑧 (𝐱)−𝐼0 (𝐱) 𝜕𝐼(𝐱)
When Δz → 0, Δ𝑧
≈ 𝜕𝑧
, and TIE is obtained. Fig. 15. The geometric illustration of the angular spectrum propagation.
2) Using Fresnel angular spectrum diffraction formula to derive
TIE in the spatial frequency domain located at the plane 𝑧 = 0 can be regarded as the linear superposition of
On the other hand, the diffraction of the coherent beam can also different plane wave components. After propagation of a distance Δz,
be described from the viewpoint of “system” in the spatial frequency the coherent field U(x, y, z) is still composed of these plane waves prop-
domain: the scalar coherent field U(x, y, 0) (assuming 𝑧 = 0) can be agating along their original directions. The amplitudes of these plane
decomposed into the coherent superposition of different angular spec- waves remain unchanged, but their optical paths traveled in each di-
trum (plane wave) components 𝑈̂ (𝑢𝑥 , 𝑢𝑦 , 0) in the 2D spatial frequency rection are different, √i.e., their phase delays are different, which can
domain. It can be presented by the following 2D Fourier transform [2] be represented as 𝑘Δ𝑧 1 − (𝜆𝑢𝑥 )2 − (𝜆𝑢𝑦 )2 . In the spatial frequency do-
∞ ∞ main, each√point in the Fourier spectrum is multiplied by an exponential
𝑈̂ (𝑢𝑥 , 𝑢𝑦 , 0)𝑒𝑗2𝜋(𝑢𝑥 𝑥+𝑢𝑦 𝑦) 𝑑 𝑢𝑥 𝑑 𝑢𝑦
2 2
𝑈 (𝑥, 𝑦, 0) =
∫ ∫
(23) term 𝑒𝑗𝑘Δ𝑧 1−(𝜆𝑢𝑥 ) −(𝜆𝑢𝑦 ) , which is equivalent to the whole 2D spectrum
−∞ −∞ multiplied by a 2D transfer function HΔz (ux , uy ). Under the precondition
∞ ∞ of 1 − (𝜆𝑢𝑥 )2 − (𝜆𝑢𝑦 )2 > 0, we get
𝑈̂ (𝑢𝑥 , 𝑢𝑦 , 0) = 𝑈 (𝑥, 𝑦, 0)𝑒−𝑗2𝜋(𝑢𝑥 𝑥+𝑢𝑦 𝑦) 𝑑 𝑥𝑑 𝑦 (24) √
∫ ∫ 𝑈̂ (𝑢𝑥 , 𝑢𝑦 , Δ𝑧) 1−(𝜆𝑢𝑥 )2 −(𝜆𝑢𝑦 )2
−∞ −∞ 𝐻Δ𝑧 (𝑢𝑥 , 𝑢𝑦 ) = = 𝑒𝑗𝑘Δ𝑧 (30)
𝑗2𝜋(𝑢𝑥 𝑥+𝑢𝑦 𝑦) 𝑈̂ (𝑢𝑥 , 𝑢𝑦 , 0)
where the exponential term 𝑒 represents a monochromatic
plane wave with unit amplitude, and the direction cosine vector of its From the perspective of linear shift invariant system, the propagation
propagation direction is (cos 𝛼, cos 𝛽), i.e., phenomenon may be regarded as a linear spatial filter with a finite
cos 𝛼 cos 𝛽 bandwidth, i.e., HΔz (ux , uy ), so so-called the angular spectrum transfer
𝑢𝑥 = , 𝑢𝑦 = (25)
𝜆 𝜆 function [2,5]. The amplitude of HΔz (ux , uy ) is zero outside a circu-
The geometric propagation of a plane wave 𝑒𝑗2𝜋(𝑢𝑥 𝑥) parallel to y-axis is lar region of radius 𝜆−1 in the frequency domain. Within that circular
illustrated in Fig. 15. The angle between the propagation direction and bandwidth, the modulus of the transfer function is unity but frequency-
x-axis is denoted as 𝛼. dependent phase shifts are introduced. It should be noted that as well as
After propagation of a distance Δz, the optical path travelled in free propagating plane waves, solutions of the wave equation are also given
space and the corresponding phase delay can be represented as by evanescent waves (from the latin evanescare, to diminish). When
√ √ 1 − (𝜆𝑢𝑥 )2 − (𝜆𝑢𝑦 )2 < 0, the evanescent waves at high spatial frequencies
Δ𝑑 = Δ𝑧 sin 𝛼 = Δ𝑧 1 − (cos 𝛼)2 = Δ𝑧 1 − (𝜆𝑢𝑥 )2 (26) diminishes exponentially with the increase of z. It decays to zero after
the propagation of a distance about several wavelengths [2,5]. In most
√
applications of optical imaging, evanescent waves have a very limited
𝜙 = 𝑘Δ𝑑 = 𝑘Δ𝑧 1 − (𝜆𝑢𝑥 )2 (27) range and can therefore be neglected. It should be mentioned the re-
Therefore, the plane wave after the propagation of a distance Δz is markable fact that, Eq. (30) is proven to be the Fourier transform of
𝑒𝑗2𝜋(𝑢𝑥 𝑥) 𝑒𝑗𝜙 . It can be seen from the Fourier transform that the point in the impulse response function [Eq. (13)] with the RS-I type obliquity
the Fourier spectrum corresponding to this plane wave is also multiplied factor, which suggests that the angular spectrum approach and the first
by an exponential term, i.e., Rayleigh-Sommerfeld integral are essentially identical in predictions of
diffracted fields [2,5].
𝑈̂ (𝑢𝑥 , 0, Δ𝑧) = 𝑈̂ (𝑢𝑥 , 0, 0)𝑒𝑗𝜙 Under the paraxial approximation, the higher order terms in the Tay-
√
lor expansion of the phase factor in Eq. (30) can be neglected
= 𝑈̂ (𝑢𝑥 , 0, 0)𝑒𝑗𝑘Δ𝑧 1−(𝜆𝑢𝑥 )2
(28)
√
Generally, for plane waves in arbitrary directions, we can get
1 − (𝜆𝑢𝑥 )2 − (𝜆𝑢𝑦 )2
√
2
𝑈̂ (𝑢𝑥 , 𝑢𝑦 , Δ𝑧) = 𝑈̂ (𝑢𝑥 , 𝑣𝑦 , 0)𝑒𝑗𝑘Δ𝑧 1−(𝜆𝑢𝑥 ) −(𝜆𝑢𝑦 )
2
(29) ( ) ( )2
1 1
= 1 − 𝜆2 𝑢2𝑥 + 𝑢2𝑦 + 𝜆4 𝑢2𝑥 + 𝑢2𝑦 + ...
2 8
To sum up, here we reveal the geometric significance of the angu- ( )
1
lar spectrum propagation: the scalar coherent optical field U(x, y, 0) ≈ 1 − 𝜆2 𝑢2𝑥 + 𝑢2𝑦 (31)
2
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

So we can get the angular spectrum transfer function under Fresnel

(paraxial) approximation
[ ( )]
𝐹
𝐻Δ𝑧 (𝑢𝑥 , 𝑢𝑦 ) = exp (𝑗𝑘Δ𝑧) exp −𝑗𝜋𝜆Δ𝑧 𝑢2𝑥 + 𝑢2𝑦 (32)

Similarly, it is not difficult to prove that Eq. (32) is the Fourier trans-
form of the Fresnel impulse response function [Eq. (17)]. In Eq. (32),
the first phase factor exp (jkΔz) is only related to the propagation dis-
tance Δz, so it can be neglected because it has no effect on the intensity
and complex amplitude distribution. Furthermore, when the propaga-
tion distance Δz → 0, the Fresnel angular spectrum transfer function can
Fig. 16. The physical implication of TIE.
be further rewritten as
𝐹 2
𝐻Δ𝑧 (𝐮) = 𝑒−𝑗𝜋𝜆Δ𝑧|𝐮| Under the paraxial approximation, the time-averaged Poynting vector
(𝜋𝜆Δ𝑧)2 |𝐮|4 can be explicitly decomposed into the transverse component and the
= 1 − 𝑗𝜋𝜆Δ𝑧|𝐮| − 2
− ...
2 longitudinal component
Δ𝑧→0 1 [ ]
≈ 1 − 𝑗𝜋𝜆Δ𝑧|𝐮|2 (33) ⟨𝐒⟩ = 𝐼(𝐱, 𝑧) ∇𝜙(𝐱)+𝑘𝐳0 (39)
𝑘
where u is the spatial frequency vector (ux , uy ). Considering the com- where 𝐳0 represents the unit vector along the z-direction. Substituting it
plex amplitude of an optical field to be recovered as 𝑈 (𝐱) = 𝑎(𝐱)𝑒𝑗𝜙(𝐱) , into Eq. (37), we can get TIE
its Fourier transform is 𝑈̂ (𝐮) = ℱ [𝑈 (𝐱)]. So the intensity at a distance 𝜕𝐼(𝐱, 𝑧)
Δz can be expressed as −𝑘 = ∇ ⋅ [𝐼(𝐱, 𝑧)∇𝜙(𝐱)] (40)
𝜕𝑧
| −1 ̂ |2 It should be noted that the time-averaged Poynting vector in Eq. (32) is
𝐹
𝐼(𝐱, Δ𝑧) = |ℱ {𝑈 (𝐮)𝐻Δ𝑧 (𝐮)}| well defined for both coherent fields and partially coherent fields.
| |
Δ𝑧→0 | −1 ̂ |2 Paganin and Nugent [159] use this formulation as the definition of
≈ |ℱ {𝑈 (𝐮)(1 − 𝑗𝜋𝜆Δ𝑧|𝐮|2 )}|
| | “phase”, which is a scalar potential of the time-averaged Poynting vec-
| [ 2 ]|2 tor. If the wave is coherent, then this is equivalent to the conventional
= −1 ̂
|𝑈 (𝐱) − 𝑗𝜋𝜆Δ𝑧ℱ |𝐮| 𝑈 (𝐮) | (34)
| | phase.
Based on the differential properties of Fourier transform, the multiplica-
2.3. The physical implication of TIE
tive frequency vector u in the Fourier domain is equivalent to the gra-
dient operator in the spatial domain, i.e., j2𝜋u → ∇. So the Laplacian
Before entering the next section about the solution of TIE, let us
operator can be represented as Δ = ∇2 → −4𝜋 2 |𝐮|2 . Substituting it into
first examine this beautiful equation. Viewed as a whole, TIE is essen-
Eq. (34), we can get
tially an expression of the energy conservation law. Under the paraxial
| 𝑗𝜋𝜆Δ𝑧 2 |2 approximation, the longitudinal energy flux of the optical field can be
𝐼(𝐱, Δ𝑧) = ||𝑈 (𝐱) + ∇ 𝑈 (𝐱)|| approximated by the intensity. Transverse energy flux is determined by
| 4𝜋 2 |
Δ𝑧 I∇𝜙, and I∇𝜙 represents the transverse component of the time-averaged
= 𝐼(𝐱) − ∇ ⋅ [𝐼(𝐱)∇𝜙(𝐱)] (35) Poynting vector, as shown in Eq. (39). Its divergence ∇ · (I∇𝜙) represents
𝑘
for the total of energy variation in the transverse plane. Longitudinal and
𝐼 (𝐱,Δ𝑧)−𝐼 (𝐱) 𝜕𝐼(𝐱)
When the propagation distance Δz → 0, Δ𝑧
≈ 𝜕𝑧
, and TIE is transverse energy dissipation should always be equal according to the
obtained. energy conservation law (as shown in Fig. 16). Examining both sides of
TIE, the left hand side (LHS) is the axial intensity derivative. In the sim-
2.2.3. Derivation of TIE based on Poynting theorem plest situation, it can be estimated by the finite difference between two
Poynting theorem [4] is in the nature a statement of conservation intensity images collected at different propagation distances along the
of energy for the electromagnetic field. In physics, the Poynting vector optical axis. More details about the axial intensity derivative estimation
represents the directional energy flux of an electromagnetic field, and will be given in Section 5. Expanding the right hand side (RHS) of TIE,
Poynting theorem states that the total dissipated energy flow through we obtain
a closed surface s of a bounded volume V in the electromagnetic field 𝜕𝐼
−𝑘 = ∇ ⋅ (𝐼∇𝜙) = ∇𝐼 ⋅ ∇𝜙 + 𝐼 ∇2 𝜙 (41)
is equal to the change over time of the complete electromagnetic en- 𝜕𝑧
ergy inside V. Since energy may be neither created nor destroyed in free where the two terms contain the phase gradient (slope, the first deriva-
space, the time-averaged Poynting vector S must obey a conservation tive) term and phase curvature (the second derivative), respectively. Ax-
equation expressed as ial variation of intensity is determined by both phase slope and phase
curvature. Phase slope induces the intensity translation, just like a prism,
⟨𝐒⟩𝑑𝑠 = 0 (36)
∮𝑠 while phase curvature induces intensity convergence or divergence, just
like a lens. So these two terms are also called prism term (slope and gra-
Its corresponding differential form is
dient) and lens term (curvature) [143] (as illustrated in Fig. 16).
∇ ⋅ ⟨𝐒⟩ = 0 (37) At last, let us consider some special cases of TIE: if the phase is con-
stant, the RHS of TIE is equal to 0, so there will be no transport of
where ∇ is the Hamiltonian in 3D space, ⟨S⟩ is the vectorial representa-
intensity effect. It is in accordance with the plane wave case discussed
tion of time-averaged Poynting vector in 3D space. The time-averaged
above. Another more meaningful situation is when the in-focus intensity
Poynting vector corresponds to the intensity (irradiance in definitions of
is uniform I ≈ constant, corresponding to the situation of a pure phase
radiometry) or flux density I of a surface element perpendicular to the
object, TIE can be simplified as
Poynting vector. When the vector potentials of electromagnetism is ne-
𝜕𝐼
glected (in the absence of phase vortices), the time-averaged Poynting −𝑘 = 𝐼 ∇2 𝜙 (42)
vector can be written in the form [159] 𝜕𝑧
This is a standard Poisson equation, revealing that a weak defocusing of
1
⟨𝐒⟩ = 𝐼(𝐱, 𝑧)∇𝜙(𝐱, 𝑧) (38) a pure phase object produces intensity contrast that is proportional to
𝑘
C. Zuo, J. Li and J. Sun et al.
Fig. 17. Three typical boundary conditions used in TIE solvers.
the phase Laplacian (curvature). This is exactly the physical implication
of the light-and-dark network structure at the bottom of the swimming
pool in Fig. 13.
3. Solutions to TIE
In this section, we will discuss how to solve TIE, which is the funda-
mental premise for the success of TIE phase retrieval. In TIE, the inten-
sity I on the RHS can be directly measured, and the LHS is the intensity
derivative along the optical axis 𝜕 I/𝜕 z which cannot be directly mea-
sured. Fortunately, it can be approximated by finite differences with a
minimum of two defocused intensity images taken at distinct planes sep-
arated by a small distance orthogonal to the optical axis (discussed in
Section 5). Thus, the only one unknown quantity in TIE is the phase func-
tion 𝜙. When the intensity I and intensity derivative 𝜕 I/𝜕 z are known,
solving the phase function is a typical inverse problem.
3.1. Boundary conditions of TIE
There are two prerequisites for uniquely determining a solution
to a partial differential equation: the partial differential equation and
boundary conditions [222]. Thus, the TIE phase retrieval is essentially a
boundary value problem: seeking a solution to TIE that also satisfies the
given boundary conditions. We then adopt a more rigorous mathemati-
cal description to express this problem: we assume the region governed
by TIE to be a general open and bounded domain Ω ⊂ ℝ2 with a piece-
wise smooth boundary 𝜕Ω. The intensity distribution I is a continuous,
nonnegative function defined on the enclosure Ω ̄ (including the region
inside Ω and the boundary 𝜕Ω), and is continuous and strictly positive
in Ω. The axial intensity derivative 𝜕 I/𝜕 z is assumed to be a continuous
function in Ω. The phase 𝜙 is expected to be single-valued and smooth in
Ω̄ (in the absence of phase vortices). In the following part, we shall con-
sider three classes of possible boundary conditions that could be applied
to solve TIE (see Fig. 17):
1) Dirichlet boundary conditions
The values of phase function 𝜙 are specified on the domain boundary
𝜙|𝜕Ω = 𝑔 (43)
Here g is a smooth function on the boundary 𝜕Ω.
2) Neumann boundary conditions
The product of the normal derivative of 𝜙 is specified on the domain
boundary
𝜕𝜙 ||
𝐼 =𝑔 (44)
𝜕𝑛 ||𝜕Ω
Optics and Lasers in Engineering 135 (2020) 106187
Here g is a smooth function on the boundary 𝜕 Ω, and 𝜕 𝜙/𝜕 n is the out-
ward normal derivative.
3) Periodic boundary conditions
The phase at the boundary repeats cyclically, e.g., for a rectangular
domain, the phase at the right boundary is the same as the phase at the
left boundary.
3.2. Well-posedness and uniqueness of the solution
The “well-posedness” and the “uniqueness” of the solution are two
essential problems of the TIE phase retrieval. The well-posedness means
that the equation must exist at least one solution to ensure the solvability
of the problem. The uniqueness of the solution means that the solution
of the equation must be unique. Obviously, the phase of the object to
be measured is well and uniquely defined. If the solution to TIE is not
unique, it will be impossible to determine whether the obtained solution
corresponds to the true phase of the object. In 1995, the well-posedness
and uniqueness of the TIE solution were first proven by Gureyev et al.
[223]: with the determined intensity I, axial intensity derivative 𝜕 I/𝜕 z,
and boundary conditions specified on 𝜕Ω, the phase can be uniquely
determined by solving the corresponding boundary value problem. For
the case of Dirichlet boundary conditions, the solution to TIE always
exists and is unique [222]. The case of Neumann boundary conditions
demands special attention because a solution may or may not exist (de-
pending on whether the compatibility condition [Eq. (45)] is satisfied
[223,224]). The compatibility condition can be obtained by integrating
both sides of TIE over the region Ω and then applying the divergence
theorem
𝜕𝜙(𝐱) 𝜕𝐼 (𝐱)
𝐼 (𝐱 ) d𝑠 = −𝑘 d𝐱 (45)
∮𝜕Ω 𝜕𝑛 ∬Ω 𝜕𝑧
where s is a parameterization of 𝜕Ω, and 𝜕I/𝜕z can be obtained by actual
measurement. Thus, the RHS of Eq. (45) is known a priori, since the func-
tion 𝜕 I/𝜕 z can be experimentally obtained, and the LHS is determined
solely by Neumann boundary conditions [Eq. (44)]. Mathematically, all
these possible Neumann boundary conditions must satisfy Eq. (45) in
order for there to be a nontrivial solution to TIE. In addition, notice that
if 𝜙 is a solution to the Neumann boundary problem, then 𝜙 + 𝐶 is also
a solution for an arbitrary constant C. In other words, the solution of
TIE subject to Neumann boundary conditions should be unique up to an
arbitrary additive constant, assuming that the solution exists in the first
place. This constant is not essential for the phase retrieval problem. The
proof of the uniqueness theorem is straightforward: we can first assume
that two qualified solutions have been exhibited and then prove they
can only differ from one additive constant using the maximum princi-
ple for elliptic equations [225]. In general, the periodic boundary con-
dition is widely adopted in the fields of lattice vibrations and electron
C. Zuo, J. Li and J. Sun et al.
Fig. 18. Phase retrieval simulations for different types of objects. (a) An isolated object located in the central field of view (FOV) (the FFT-based solver gives accurate
reconstruction); (b) a complex object extending outside the image boundary (the FFT-based solver produces large boundary artifacts); and (c) the DCT solver with a
hard aperture (the inhomogeneous boundary conditions can be measured at the boundary, which produces accurate phase reconstruction even if the object is located
at the aperture boundary).
theory of metals, wherein most objects of interest are characterized by
periodicity. In most cases of QPI, the samples are not periodically dis-
tributed, so periodic boundary conditions are difficult to meet. The pur-
pose of introducing periodic boundary conditions into the solution of
TIE is to simplify the numerical implementation, which will be detailed
in Subsection 3.3. For periodic boundary conditions, the solution to TIE
is also unique up to an arbitrary additive constant.
In summary, regarding the well-posedness and uniqueness of the so-
lution to TIE, the following conclusions can be drawn:
1) For Dirichlet boundary conditions, the solution to TIE always exists
and is unique.
2) For Neumann boundary conditions, the solution to TIE may or
may not exist, depending on whether the compatibility condition of
[Eq. (45)] is satisfied. When the compatibility condition is satisfied,
the solution to TIE exists and is unique up to an arbitrary additive
constant.
3) For periodic boundary conditions, the solution to TIE always exists
and is unique up to an arbitrary additive constant.
3.3. Compatibility condition and energy conservation law
It is very interesting to examine the physical picture described by the
compatibility condition [Eq. (45)]: actually, it can be thought as an ex-
pression of local energy conservation law in bounded domains: the loss
of energy (intensity) inside the domain arising from energy flow across
the boundary. If the boundary condition does not satisfy Eq. (45), it is
not physically grounded so that the correct solution cannot be obtained.
If we extend the bounded domain Ω to unbounded free space, the con-
tour integral vanishes and the compatibility condition becomes
𝜕𝐼 (𝐱)
d𝐱 = 0
∬ℝ 2 𝜕𝑧
It represents the law of energy conservation for unlimited free space. In
essence, the energy conservation expressed by Eqs. (45) and (46) is a
universal law of physics, so it imposes an implicit physical constraint on
the solution of TIE.
Optics and Lasers in Engineering 135 (2020) 106187
3.4. Solutions to TIE
The essence of TIE retrieval is to solve the partial differential equa-
tion under appropriate boundary conditions, and how to solve TIE ac-
curately and efficiently has become one of the most popular and chal-
lenging research directions since its inception. In the paper of Teague
[141] in 1983, he tried to simplify TIE into standard Poisson equations
by introducing an auxiliary function. This auxiliary function 𝜓 is so-
called “Teague’s auxiliary function”, which satisfies
𝐼 (𝐱)∇𝜙(𝐱) = ∇𝜓 (𝐱) (47)
The physical meaning of this auxiliary function has been discussed in
Subsection 2.2. I∇𝜙 can be interpreted as the time-averaged transverse
Poynting vector [159] and can be characterized by the scalar potential
𝜓. With this auxiliary function 𝜓, TIE can be simplified into two stan-
dard Poisson equations [76,141,201]
𝜕𝐼 (𝐱) 1
= − ∇ 2 𝜓 (𝐱 ) (48)
𝜕𝑧 𝑘
[ ]
∇ ⋅ 𝐼 (𝐱)−1 ∇𝜓 (𝐱) = ∇2 𝜙(𝐱) (49)
The Teague’s auxiliary function 𝜓 can be solved from Eq. (48), and
then substituted into Eq. (49), the phase function can be uniquely de-
termined by phase integration (up to an arbitrary additive constant). In
order to solve these two Poisson equations, Teague [141] introduced the
Green’s function method to derive the first analytical solution to TIE, al-
though, to our knowledge, this has never been implemented. After that,
various numerical TIE solvers have been proposed, such as the multi-
grid method [76,226–228], the Zernike polynomial expansion method
[156,229,230], the FFT-based method [76,149,158,159,231,232], and
the discrete cosine transform (DCT) method [201–203]. Multi-grid
method is a common method to solve partial differential equations,
which is mainly used for the rectangular domain and relatively com-
plicated to implement [76,226–228]. Since Zernike polynomials are or-
(46) thogonal functions defined on the unit circle, the Zernike polynomial
expansion method is used primarily in wavefront sensing and optical
aberration characterization [156,229,230]. The FFT-based method is
probably the most popular TIE solvers especially in the field of QPI for
its simplicity, efficiency, and applicability for the case of non-uniform
intensity distribution. The present widely used version of the FFT solver
C. Zuo, J. Li and J. Sun et al.
is developed by Paganin and Nugent [159], which can be considered
a generalization of the Gureyev and Nugent’s approach [158] as it re-
duces to that form in the case of nonuniform intensity. Under general
nonuniform intensity, it takes the following form [159]
[ ]
𝜕𝐼 (𝐱)
𝜙(𝐱) = −𝑘∇−2 ∇ ⋅ 𝐼 −1 (𝐱)∇∇−2 (50)
𝜕𝑧
where ∇−2 is the inverse Laplacian operator, denoting a simplified rep-
resentation of the solution to the corresponding Poisson equation. Note
that the inverse Laplacian operator is effectively implemented with
FFT based on differential properties of the Fourier transform (∇−2 →
− 21 2 ) [76,159,232].
4𝜋 |𝐮|
3.4.1. Solving TIE without boundary values
Despite the mathematical well-posedness, the rigorous implementa-
tion of TIE phase retrieval tends to be difficult because the associated
boundary conditions are difficult to measure or to know as a priori. For
example, in Teague’s Green function solution [141], one needs to know
the phase value at the region boundary as the Dirichlet boundary con-
dition, which is lack of feasibility because the phase function is exactly
what we want to measure. Generally speaking, it is difficult to obtain any
prior information about the measured object before taking any measure-
ment (Teague recommends that the boundary value can be measured by
using a Shack-Hartmann wavefront sensor, which is indeed very incon-
venient for practical implementation). Parvizi et al. [233] pointed out
that the Dirichlet boundary condition can be obtained by manually se-
lecting the “smooth region” inside the FOV (the region without the ob-
ject is considered ‘flat’ and has a constant phase value). However, this
method also needs the prior information about the object and requires
human intervention. The FFT-based method does not need to impose
the boundary condition explicitly, but it assumes that the finite field is
periodic and repetitive. This implies periodic boundary conditions due
to the cyclic nature of the discrete Fourier transform. However, this sit-
uation is rather restrictive and does not reflect general experimental
conditions. When the actual experimental situation violates those im-
posed assumptions, e.g., objects are located at the image borders, severe
boundary artifacts will appear, seriously affecting the accuracy of the
phase reconstruction [201,202,234].
Because the boundary signals are not easy to acquire experimen-
tally, some researchers sought new ways to solve TIE without taking any
boundary measurements. Coincidentally, all these efforts aim to nullify
the LHS of Eq. (45), making boundary conditions unnecessary
𝜕𝜙(𝐱) 𝜕𝐼 (𝐱)
𝐼 (𝐱 ) d𝑠 = 0 𝑜𝑟 equivalently d𝐱 = 0 (51)
∮𝜕Ω 𝜕𝑛 ∬Ω 𝜕𝑧
This is actually equivalent to applying the energy conservation law for
unbounded space to a bounded domain directly. Generally, Eq. (51) does
not hold for a bounded region unless the contour integral [LHS of
Eq. (45)] vanishes. Physically, it is equivalent to the special case that
there is no overall energy transfer through the domain boundary. To
solve TIE without boundary values, there are three common ways to
satisfy Eq. (51):
1) “Soft-edged” illumination
The soft-edged illumination suggested by Gureyev and Nugent
[157] represents a special case that the intensity is smooth and strictly
greater than 0 inside the domain (I > 0) but strictly vanishes at the
boundary 𝐼𝜕Ω = 0. In this case, they proven that the solution to TIE ex-
ists and is unique to an arbitrary additive constant. Although Gureyev
and Nugent [157] called this special case as a kind of “unconventional”
boundary condition, by comparing it with Eq. (44), it can be found that it
just corresponds to a special case of the homogeneous Neumann bound-
ary conditions ( 𝐼 𝜕𝜙∕𝜕𝑛|𝜕Ω = 0). The disadvantage of this method is that
defining the boundary of the domain tends to be very difficult in prac-
tice since “very-near zero” and “at zero” may seem almost the same but
can be entirely different vis-a-vis the well-posed TIE.
Optics and Lasers in Engineering 135 (2020) 106187
2) “Flat” boundary
This is the simplest and the most common way to bypass the diffi-
culty associated with the boundary condition problem. One can place
the sample carefully so that it is centrally isolated in the image FOV,
surrounded by an unperturbed plane wave, so that the boundary of
the image is “empty” (sample-free). Then the phase at the boundary
can be considered as “flat”, and zero phase changes at the boundary
(corresponding to homogeneous Neumann boundary conditions) can be
assumed thus the Eq. (51) can be satisfied. In fact, in this case one
can define not only homogeneous Neumann boundary conditions (con-
stant phase 𝜙|𝜕Ω = 𝐶 at the boundary), but any homogeneous Dirichlet
boundary conditions ( 𝐼 𝜕𝜙∕𝜕𝑛|𝜕Ω = 0) or even periodic ones (the phase
at the boundary repeats cyclically). In such a case, the phase can be re-
covered accurately regardless of the selected boundary conditions, for
example, using the FFT-based method with periodic boundary condi-
tions. However, this configuration does not reflect general experimental
conditions, and is impractical when the object is larger than the cam-
era FOV. This kind of boundary error is problematic when dealing with
a extended object that must cover the image boundary, e.g., microlens
arrays [201,202,234].
3) Mirror padding
This method is proposed by Volkov et al. [231] to solve TIE with-
out boundary value measurements. It has no additional requirements
about the test object and experimental conditions, but only needs to
symmetrically extend the intensity images into a four times larger size
so that the homogeneous Neumann boundary conditions can be satisfied
( 𝜕𝜙∕𝜕𝑛|𝜕Ω = 0), and the LHS of Eq. (45) vanishes. Although Volkov et al.
[231] demonstrated that this method performs better than the tradi-
tional FFT-based method in certain circumstances, it is only a pure math-
ematical trick to nullify the energy flow across the boundary through
appropriate symmetrization of input images. It assumes there is no en-
ergy dissipation through the image boundary, which generally does not
satisfy the compatibility condition [Eq. (45)], and thus, is not physically
grounded. Note severe boundary errors can be observed when objects
are located at the image boundary. Such kind of artifacts are not just
limited to the boundary region but may propagate inside the domain
and degrade the reconstruction accuracy prevailingly [201,202].
3.4.2. Solving TIE with boundary value measurements
The simplified solutions to TIE discussed in Subsection 3.4.1 can
solve TIE without boundary value measurements, but it does not mean
that we can solve TIE without imposing any boundary conditions. To be
more specific, we have to restrict the test objects or experimental con-
ditions to certain “implicit boundary conditions”. On the other hand,
for certain phase functions (such as tilt, defocus, astigmatism, and other
phase aberrations with zero or constant curvature), the defocus-induced
intensity derivative signals are all concentrated at the boundary region.
If the boundary conditions are not considered, the phase can never be
recovered [235–237]. The only solution to issues mentioned above is
to solve TIE with correct boundary conditions, and the key problem is
how to obtain the boundary signals experimentally. Roddier’s research
group [145,146] first explored this problem and found that the bound-
ary values required for the simplified version of TIE [Eq. (42)] with
uniform intensity can be obtained from the intensity measurements at
the pupil boundary. Because the boundary signals can be acquired ex-
perimentally, Roddier et al. [145,146] used the SOR and the iterative
Fourier transform [149] to solve the simplified TIE effectively. Later this
method was adopted by Woods and Greenaway [238], and the Green’s
function method was used to solve the simplified TIE. Although Rod-
dier’s method addresses the problem of boundary value measurements,
it is still necessary to distinguish the boundary signals from the interior
intensity derivative, which has been known to cause serious difficulties
[147,156,235,239]. Moreover, it relies on one major assumption: the in-
tensity distribution should be uniform within the domain (I ≈ constant).
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

This may be a reasonable assumption in the field of adaptive optics, but

it is difficult to satisfy in the field of QPI, especially for objects with
strong absorptions.
To solve the above-mentioned problems, Zuo et al. [201] found that
even if the intensity distribution is non-uniform, the inhomogeneous
Neumann boundary conditions can still be obtained when a hard-edged
̄ 𝐴Ω = 0 when 𝐫 ∉ Ω)
aperture (the aperture function 𝐴Ω = 1 when 𝐫 ∈ Ω, ̄
is introduced at the object plane (𝑧 = 0 or its conjugated plane). Sub-
stituting the aperture function into TIE, the intensity transport can be
written in the following form
𝜕𝐼 ( ) 𝜕𝜙
−𝑘 = 𝐴Ω 𝐼 ∇2 𝜙 + ∇𝐼 ⋅ ∇𝜙 − 𝐼 𝛿 (52)
𝜕𝑧 𝜕𝑛 𝜕Ω
where 𝛿 𝜕Ω is the Dirac delta function around the aperture edge.
Eq. (52) suggests that the axial intensity derivative signals consist of Fig. 19. Helmholtz decomposition of the transverse flux field. The x and y com-
two non-overlapping components: ponents of the vector fields are shown in the first row and the second row, re-
spectively. The term ∇ × 𝜂 is missing in Teague’s assumption.
1) The intensity variation inside the domain due to the phase slope and
curvature as if the aperture is not present.
2) A delta-function-like signal sharply peaked at the aperture bound- 2) The effectiveness of the method relies on the hard-edged aperture.
ary, which provides the exact Neumann boundary conditions for TIE For the case of “soft-edged” aperture such as a Gaussian beam, al-
[Eq. (44)]. though the energy conservation law can be satisfied, the algorithm
is still not applicable in this case.
Since the whole axial intensity derivative [the LHS of Eq. (52)] is
experimentally measurable through finite-difference scheme (note the To address these problems, Huang et al. [203] further extended the
aperture should be smaller than the image FOV so that all the bound- DCT solver to an arbitrarily shaped aperture by an iterative compensa-
ary signals can be captured), and the two RHS terms do not overlap in tion mechanism. This method takes the DCT-based solution as the initial
space, there is enough information to solve TIE uniquely without requir- value, and then it is refined by the iterative compensation algorithm to
ing prior knowledge of the boundary conditions. The inhomogeneous obtain an accurate solution. Thus, the DCT-based method can be ex-
boundary value problem has been proven to be well-posed because it tended to an arbitrarily shaped aperture. Ishizuka et al. [240,241] suc-
automatically satisfies the following compatibility condition [derived cessfully applied the iterative DCT solver to recover the additional phase
from integrating both sides of Eq. (52) over the whole space] term corresponding to the curvature of field on the image plane in
TEM. A similar iterative TIE solver based on FFT was also proposed by
𝜕𝐼 (𝐱) 𝜕𝐼 (𝐱) 𝜕𝜙(𝐱)
−𝑘 d𝐱 = −𝑘 𝑑𝐱 − 𝐼 (𝐱) d𝑠 = 0 (53) Mehrabkhani et al. [242].
∬Ω̄ 𝜕𝑧 ∬Ω 𝜕𝑧 ∮𝜕Ω 𝜕𝑛
This is the same expression as the compatibility condition of the Neu- 3.5. Phase discrepancy analysis and compensation
mann boundary problem [Eq. (45)], which leads to the inescapable con-
clusion that the solution to this Neumann boundary problem always ex- Another important issue regarding the solution of TIE is “phase dis-
ists and is unique up to an arbitrary additive constant. On this basis, crepancy”, which represents the difference between the ground truth
Zuo et al. [201] developed the DCT-based TIE solver using experimen- phase and the phase retrieval by TIE. This is because most of TIE
tally measurable boundary conditions. The solution takes the following solvers, such as FFT and DCT, introduce the Teague’s auxiliary func-
form tion [141] [Eq. (47)] to simplify TIE into standard Poisson equations.
[ ]
−2 𝜕𝐼 (𝐱 ) The Teague’s assumption suggests that the transverse flux is conserva-
𝜙(𝐱) = −𝑘∇−2𝐷𝐶𝑇 ∇ 𝐷𝐶𝑇 ⋅ 𝐼 −1
( 𝐱 )∇ 𝐷𝐶𝑇 ∇ 𝐷𝐶𝑇 𝜕𝑧 . (54)
tive so that a scalar potential 𝜓 exists and satisfies Eq. (48). However,
it is important to remark that the Teague’s auxiliary function does not
Eq. (54) is very similar to the FFT-based method proposed by Paganin
always exist in practical situations since the transverse energy flux may
and Nugent [Eq. (50)] [158,159], and the only difference is that the
not be conservative [I(x)∇𝜙(x) is an ordinary 2D scalar field]. Accord-
inverse Laplacian operator ∇−2 𝐷𝐶𝑇
and the gradient operator ∇DCT are
ing to the Helmholtz’s theorem, the transverse flux can be decomposed
implemented by the DCT rather than FFT (because the eigenfunctions
in terms of the gradient of a scalar potential 𝜓 and the curl of a vector
of the Green’s function associated with this Neumann boundary value
potential 𝜂
problem is the Fourier cosine harmonics, rather than the general Fourier
exponential harmonics). In addition, it should be noted that all the 𝐼 (𝐱)∇𝜙(𝐱) = ∇𝜓 (𝐱) + ∇ × 𝜂(𝐱) (55)
source data and related computations must be strictly defined on the
closed rectangular region Ω,̄ which includes both the aperture boundary With certain (Dirichlet, Neumann, or periodic) boundary conditions, this
and the region inside it (so that all the boundary signals can be enclosed decomposition is unique (or unique up to a vectorial constant that may
̄ In this way, the measured intensity derivative can be treated as
in Ω). float between the two components). Compared with Eq. (47), it is plain
one entity without requiring special-purpose detection schemes to ex- to see that the term ∇ × 𝜂(x) is ignored in Teague’s assumption, making
plicitly extract the boundary signals. Therefore, this approach solves a silent hypothesis that the transverse flux I(x)∇𝜙(x) is irrotational (see
a series of problems associated with the generation, acquisition, and Fig. 19). However, it is important to remark that the Teague’s auxiliary
separation of boundary signals in actual experimental conditions. The function does not always exist in practical situations since the transverse
effectiveness of this method has been experimentally verified by the energy flux may not be conservative, and consequently it would produce
characterization of micro-optical elements [202]. Despite its success in results that would not adequately match the exact solution.
solving TIE under inhomogeneous Neumann boundary conditions, the In fact, Allen and Oxley [76] had noticed this phase discrepancy
DCT-based solver still has two limitations: problem as early as 2001. Ten years later, Schmalz et al. [244] made
a detailed theoretical analysis on the phase discrepancy problem based
1) The fast DCT algorithm is limited to the rectangular region, and thus, on Helmholtz decomposition theorem. They also provided a simulation
the DCT-based solver cannot be directly applied to non-rectangular example to demonstrate that the phase discrepancy induced by Teague’s
(such as circle, ring, etc.) or irregular regions. auxiliary function can be significant [mean square error (MSE) larger
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

than 9%]. In 2014, Zuo et al. [204] examined the effect of the missing
rotational term on phase reconstruction and derived the necessary and
sufficient condition for the validity of Teague’s assumption
∇𝐼 (𝐱)−1 × ∇−2 {∇ ⋅ [∇𝐼 (𝐱) × ∇𝜙(𝐱)]} = 0
Eq. (56) shows that if the in-focus intensity distribution is nearly uni-
form, the phase discrepancy resulting from the Teague’s auxiliary func-
tion is quite small and negligible. This is why the issue of “phase dif-
ferences” has not received enough attention before. However, when the
measured sample exhibits strong absorption, the phase discrepancy may
be relatively large and cannot be neglected, just as shown in the paper
of Schmalz et al. [244]. To compensate the phase discrepancy owing to
Teague’s assumption, Zuo et al. [204] further developed a Picard-type
iterative algorithm, in which the phase is gradually accumulated until
a self-consistent solution is obtained. A similar iterative compensation
method was also proposed by Shanker et al. [245]. Within two to four
iterations, the phase discrepancy can be reduced to a negligible level,
and the exact solution to TIE can be thus obtained.
At the end of this section, we can conclude that there are at least
four issues need to be addressed for a desired TIE solver:
1) It should account for inhomogeneous boundary conditions with ex-
perimentally measured boundary signals.
2) It provides an accurate solution that is free from phase discrepancy.
3) For periodic boundary conditions, the solution to TIE always exists
and is unique up to an arbitrary additive constant.
4) It should be computationally efficient and strictly convergent (if it
is iterative).
A further attempt to this objective was recently reported by Zhang
et al. [243], where a universal solution to TIE is proposed with the
features of high accuracy, strict convergence guarantee, applicability
to arbitrarily-shaped regions, and simplified implementation and com-
putation. This method is formally similar to previous iterative solvers
[203,204,242,245] except that the maximum intensity assumption is
introduced to guarantee the rigorous convergence and simplify the im-
plementation of the iterative algorithm. The comparisons of the different
TIE solvers and the associated boundary conditions are summarized in
Table 3.
To aid the reader, we have uploaded the MATLAB source codes
for 4 typical TIE solvers described in this section: the FFT-based solver
proposed by Paganin and Nugent [159], the DCT-based solver proposed
by Zuo et al. [202], the iterative DCT-based solver proposed by Huang
(56)
et al. [203], and the iterative universal solution proposed by Zhang et al.
[243]. The code is composed of 4 simulation demo files, 3 experimen-
tal demo files, and supplemental experimental data and functions, as
presented in the uploaded files.
In the simulations, the intensity measurements are designed at a
wavelength of 0.633 𝜇m, and the image is composed of 256 × 256
pixels with a pixel pitch of 2.2 𝜇m. The defocused image is taken
at a distance of 1𝜇m from the object plane. In the first simulation
(“Main_Sim1_Isolated.m”), we create an isolated object located in the
central FOV with different types of intensity distribution [e.g., uniform
intensity (Case 1), nonuniform intensity without small intensity val-
ues (Case 2), nonuniform intensity with small intensity values (Case
3), nonuniform intensity with intensity zeros (Case 4)]. In the second
simulation (“Main_Sim2_Extended.m”), the object is extended over the
FOV boundary, and the intensity distributions are same as those in the
first simulation. In the third simulation (“Main_Sim3_Wave front.m”),
we simulate a smooth wavefront defined on a circular aperture with
different types of intensity distribution [e.g., uniform intensity (Case
1), hard-edged Gaussian intensity without small intensity values (Case
2), and soft-edged Gaussian intensity with small intensity values (Case
3)]. In the fourth simulation (“Main_Sim4_Arbitrary.m”), we change the
shape of the aperture arbitrarily, including elliptical aperture (Case 2),
incomplete elliptical aperture (Case 3), annular aperture (Case 3), in-
complete annular aperture (Case 4), telescope aperture (Case 5), and
incomplete telescope aperture (Case 6).
Three experimental demo files are also provided to validate these TIE
solvers. In the first experiment (“Main_Exp1_SingleCell.m”), the sam-
ple is an individual HeLa cell located in the central FOV. In the second
experiment (“Main_Exp2_Microlens.m”), the sample is a microlens ar-
ray limited by a (imperfect) rectangular aperture. In the third experi-
ment (“Main_Exp3_MultipleCells.m”), several HeLa cells are randomly
distributed within the FOV limited by an (imperfect) octagonal aper-
ture. All experiment data are collected using an off-the-shelf wide-field
microscope with the built-in Köhler illumination. The condenser aper-
ture diaphragm of the microscope is closed down, and an interference
filter (central wavelength 550nm) is inserted in the illumination path
to produce quasi-monochro matic, spatially coherent illumination. For

Table 3
Comparison of the solutions and the boundary conditions of TIE

Issues Techniques Pros Cons

TIE solvers Green’s function [141,238] Theoretical analysis Lack of feasibility

Multi-Grid [76,227]
Zernike polynomials [156,157]
FFT [76,157–159]
DCT [201,202]
Iterative DCT [203]
Iterative Universal Solution [243]
Simple and fast Low-frequency noise
Precisely represent the optical aberration Circular domain only, difficult to follow
details
Fast, easy to implement, incorporate Imply periodic boundary conditions
regularization in reconstruction
Fast, inhomogeneous Neumann boundary Rectangular domain only, aperture required to
conditions limit the FOV
Inhomogeneous Neumann boundary Need several iterations
conditions, arbitrarily-shaped apertures
Inhomogeneous Neumann boundary Need several iterations
conditions, arbitrarily-shaped apertures, strict
convergence guarantee

Boundary conditions
Homogeneous Dirichlet/Neumann
[231,234]
Periodic [76,157–159]
Inhomogeneous Dirichlet [141]
Inhomogeneous Neumann [201–203]
Easy to apply, can be implemented by “Flat” (constant phase) at boundary
different solvers
Can be implemented by FFT-based solver Periodic phase at boundary
- Phase values at boundary required
Can be measured by introducing a hard -
aperture

Phase discrepancy Picard-type iteration [204] Can compensate the phase discrepancy Need 2-4 iterations
C. Zuo, J. Li and J. Sun et al.
Fig. 20. Schematic of a wide-field microscope based on the Köhler illumination configuration - 6f imaging system.
more detailed explanations about the MATLAB source codes, readers can
refer to the associated “Data in Brief” article entitled “MATLAB source
codes and dataset for transport of intensity equation: a tutorial”.
4. Image formation of coherent imaging and microscopy
The basic idea of the computational phase microscopy based on TIE
is to build an accurate forward image formation model and then imple-
ment the phase retrieval algorithm by solving the corresponding inverse
problem. Although TIE describes the quantitative relationship between
intensity and phase in the wave propagation, it depends on two restric-
tive assumptions: paraxial approximation and the weak defocusing. On
the other hand, the derivation of TIE, as we discussed in Subsection 2.2,
does not consider the finite aperture effect of the microscopic system.
Therefore, in this section, we will consider a more practical microscopic
imaging system and present the corresponding image formation models
under coherent illumination. In addition, the relationship between these
phase imaging models and the original TIE will be discussed.
4.1. Illumination coherence and apertures of a microscopic imaging system
A typical microscopic imaging system based on Köhler illumination
can be described by the 6f system [142,178,246], as shown in Fig. 20. In
Köhler illumination, an image of the light source is focused at the con-
denser aperture diaphragm to produce parallel light through the plane
of the specimen or object. In this imaging system, the incoherent light
source (the condenser aperture diaphragm) is collimated by the con-
denser lens to produce parallel light through the plane of the specimen
or object. Then the object is imaged by the objective lens to form the fre-
quency spectrum in the objective aperture plane. This spectrum is lim-
ited by the aperture diaphragm of the objective lens and then produces
the image of the object on the image plane. In this imaging model, axial
defocus is usually modeled as an angular spectrum transfer function ap-
plied to the objective aperture. In the following, the formal description
of the imaging is limited to unit (1.0 × ) magnification. An extension to
arbitrary magnification is straightforward.
In a microscope based on the Köhler illumination configuration,
there are two important apertures that should be considered. The first
is the objective pupil defined by the NA of the objective lens, which
is a dimensionless number that characterizes the range of light angles
over which the imaging system can accept. The NA of the objective lens
is a very important parameter because it determines the spatial resolu-
tion of the microscope imaging system. For a diffraction-limited optical
imaging system, higher NA means it can gather light with larger an-
gles, thereby having a more powerful ability to resolve fine specimen
Optics and Lasers in Engineering 135 (2020) 106187
detail. The condenser (illumination) aperture is another very important
aperture of a microscope. Generally, people often only notice that the
imaging resolution of a microscope is determined by the NA of the ob-
jective lens. In fact, the NA of the illumination has equal importance as
well [247] (more detailed discussions can be found in Section 6). The
size of the condenser aperture diaphragm can be used to control the NA
of the light cone that illuminates the sample, or equivalently, the spatial
coherence of the illumination.
Why is the concept of “coherence” so important in optical imaging?
Because the coherence is the basic characteristic of light as a wave to
form interference or diffraction phenomenon, it describes all properties
of the correlation between physical quantities of a single wave, or be-
tween several waves or wave packets. Two beams of light are coherent
when the phase difference between their waves is constant; they are
incoherent if there is a random or changing phase relationship. Stable
interference patterns are formed only by radiation emitted by coherent
sources, ordinarily produced by splitting a single beam into two or more
beams. A laser, unlike an incandescent source, can produce a beam in
which all the components bear a fixed relationship to each other. When
we discuss the propagation characteristics of light and the derivation of
TIE in Section 2, the optical field is assumed to be perfectly coherent,
i.e., the object is illuminated with a plane wave obtained by collimating
a monochromatic point source. However, in practice, the light source
may not be strictly monochromatic and is of finite extent. These two
aspects are related to the temporal coherence (spectrum) and spatial
coherence (size) of light source, respectively [5,248]. More intuitively,
temporal coherence is a measure of the correlation of light wave’s phase
at different points along the direction of propagation (longitudinal di-
rection). It tells us how monochromatic a source is. Generally, a laser
with good monochromaticity can be considered as a temporally coher-
ent source. In contrast, a broadband source is temporally incoherent,
such as the light emitted from the incandescent lamp or the sun. The
spatial coherence is a measure of the correlation of a light wave’s phase
at different points transverse to the direction of propagation. It tells us
how uniform the phase of a wavefront is and is related to the dimension
of the light source. Thus, the light emitted from an ideal point source is
spatially coherent, while the light from a highly extended source is spa-
tially incoherent. Fig. 21 gives several typical examples of light sources
with different degrees of temporal coherence and spatial coherence. In
the general concept, the sunlight is considered completely incoherent,
but the fact that it would not always be the case. The reason is that al-
though sunlight is a broadband source (temporally incoherent), it can
be approximated as a point source because the sun is far enough from
us. When the sky is cloudless (the atmospheric scattering effect can be
neglected), the sunlight reaching the earth’s surface can be considered
C. Zuo, J. Li and J. Sun et al.
Fig. 21. Several typical examples of light sources with different degrees of tem-
poral coherence and spatial coherence.
Fig. 22. Illustration of light wave coherence. An incoherent light source emits
spherically shaped wave-fronts and propagates a certain distance, then passes
through a pinhole so that the spatial coherence of light is improved. It then
passes through a narrow wavelength filter to yield quasi-monochromatic light,
achieving in this way temporal and spatial coherence.
spatially coherent. Similarly, the laser is not always completely coher-
ent. It is only temporally coherent. In some special cases, e.g., when the
laser light passes through strongly scattering medium (e.g., a rotating
ground glass), the spatial coherence of the light will completely vanish.
Let us go back to the microscopic imaging system. In the simplified
illumination path of Fig. 22, we assume the source is an extended source
with a broad spectrum (temporally and spatially incoherent). The light
emitted from the source is firstly limited by the condenser aperture di-
aphragm (a circular mask with diameter d), and then passes through
an ideal monochromatic filter (wavelength 𝜆) to form the illumination
field in the object plane. At this time, if the aperture diaphragm is closed
to the minimum (like an ideal pinhole), the illumination reaching the
object plane can be regarded as spatially coherent (perfect spherical or
plane waves). However, if the aperture diaphragm is gradually opened
up, the spatial coherence of the illumination field will decrease accord-
ingly. In a microscopic system, the spatial coherence of the illumination
is often quantified by the coherence parameter, which is defined as the
Optics and Lasers in Engineering 135 (2020) 106187
Fig. 23. Varying the illumination aperture to control the spatial coherence by
opening up the condenser diaphragm of a phase contrast microscope. (a) 𝑠 =
0.15; (b) 𝑠 = 0.3; (c) 𝑠 = 0.55; (d) 𝑠 = 0.8; (e) 𝑠 = 1.0.
ratio of the illumination NA to the objective NA [177,178,246,249]
𝑁 𝐴𝑖𝑙𝑙
𝑠= (57)
𝑁 𝐴𝑜𝑏𝑗
The ratio of the illumination NA to the objective NA is equivalent to
the radius ratio of the aperture diaphragm of the condenser to that of
the objective lens. Therefore, by inserting a Bertrand lens or a phase
telescope into the beam path to image the back focal plane of the ob-
jective lens, the coherence parameter can be estimated by measuring
the ratio of the condenser aperture diaphragm to the objective aperture
diaphragm (dashed line in Fig. 23). In fact, although s is called “coher-
ence parameter”, it represents the degree of “incoherence” of the imaging
system. Specifically, a larger s means weaker spatial coherence, so strict
spatial coherence requires s → 0, and s → ∞ refers to complete incoher-
ence. The coherence parameter of the microscopic imaging system is
closely related to image formation. In the case of incoherent imaging,
the sample’s phase component will no longer affect the intensity distri-
bution of the image. For weak phase objects, the microscope behaves
like a completely incoherent system when s ≥ 1, which means the phase
effect vanishes [178,179,249]. Therefore, 0 < s < 1 is generally consid-
ered partially coherent, which is a beneficial setting for phase imaging.
It can provide higher imaging signal-to-noise ratio (SNR) and improved
imaging resolution. The relevant content will be discussed in detail in
Section 6. In this section, we mainly discuss the image formation model
and phase retrieval problem under an ideal coherent imaging system.
4.2. Ideal imaging model under coherent illumination
In the following section, we only consider the case of coherent il-
lumination: the light source is a geometric point on the optical axis of
the condenser aperture plane, and it has strict monochromaticity. The
imaging process satisfies the ideal coherent imaging model, and can be
described as a linear system for complex amplitude. Considering a thin
object with complex transmittance 𝑇 (𝐱) = 𝑎(𝐱) exp[𝑗𝜙(𝐱)] illuminated by
a monochromatic plane wave, the complex amplitude distribution in the
object plane can be expressed as
𝑈0 (𝐱) = 𝑇 (𝐱) = 𝑎(𝐱) exp [𝑗𝜙(𝐱)] (58)
Without loss of generality, we assume that the plane wave illumina-
tion has unit intensity. The complex amplitude distribution in the image
plane Ui (x) can be represented as the convolution of the complex am-
plitude distribution of the ideal object field U0 (x) and the point spread
function (PSF) of the imaging system h(x)
𝑈 𝑖 (𝐱 ) = 𝑈 0 (𝐱 ) ∗ ℎ (𝐱 ) (59)
Thus, the captured image intensity can be expressed as
𝐼𝑖 (𝐱) = 𝑈𝑖 (𝐱)𝑈𝑖∗ (𝐱) = ||𝑈0 (𝐱) ∗ ℎ(𝐱)||
2
(60)
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Taking Fourier transform on both sides of Eq. (59), the following ex-
pression in the frequency domain can be obtained:
𝑈̂ 𝑖 (𝐮) = 𝑈̂ 0 (𝐮)𝐻 (𝐮) (61)
The Fourier transform of the intensity passing through the imaging sys-
tem can be expressed as
{ }
| [ ]|2
𝐼̂𝑖 (𝐮) = 𝑈̂ 𝑖 (𝐮) ⊗ 𝑈̂ 𝑖∗ (𝐮) = ℱ |ℱ −1 𝑈̂ 0 (𝐮)𝐻 (𝐮) | (62)
| |

where 𝑈̂0 (𝐮) and 𝑈̂ 𝑖 (𝐮) are the Fourier transforms of the complex
amplitudes U0 (u) and Ui (u), respectively, and 𝐻 (𝐮) = ℱ {ℎ(𝐱)} is the
transfer function under coherent illumination. For a diffraction-limited
aberration-free imaging system, the coherent transfer function is just
the pupil function of the objective 𝐻 (𝐮) = 𝑃 (𝐮), which is a standard circ
function with unit amplitude across the passband (without attenuation)
[2,5]. This indicates that we cannot get an ideal point when imaging of Fig. 24. The first four orders of Bessel function of the first kind.
an ideal point object even with a perfect aberration-free lens, instead,
we can only obtain a diffraction spot determined by the limited lens
aperture. In addition, when the sample is located at a defocused plane
with a defocus distance Δz, the coherent transfer function is determined
by the product of the pupil function and the angular spectrum transfer
function of the objective lens [Eq. (30)]
√
2
𝐻 (𝐮) = 𝑃 (𝐮)𝐻Δ𝑧 (𝐮) = 𝑃 (𝐮)𝑒𝑗𝑘Δ𝑧 1−𝜆 |𝐮| ,
2

( ) {
𝐮 1 |𝐮| ≤ 𝑁𝐴
𝑃 (𝐮) = 𝑐 𝑖𝑟𝑐 = 𝜆 (63)
𝑁𝐴∕𝜆 0 else

where NA denotes the numerical aperture of the objective lens. Com-

pared with the pupil function, this expression includes one addi-
tional phase factor to account for the defocusing effect, thus this co-
herent transfer function is often referred to as the defocused pupil
function. There are two points need to be further elaborated here:
first, the[ coherent
(√ transfer function
)] is sometimes written as 𝐻 (𝐮) =
𝑃 (𝐮)exp 𝑗𝑘Δ𝑧 1 − 𝜆2 |𝐮|2 − 1 , in order to guarantee that it can be
simplified into a form without the z-dependence [the term of exp (jkΔz)].
Under the paraxial approximation without affecting the transverse in-
Fig. 25. Intensity distribution of Airy disk. (a) 2D intensity distribution of Airy
tensity and phase distribution calculated from Eq. (32). In addition,
disk; (b) cross-sections of Airy disk intensity and Airy function; (c) the zero-
microscopic systems are generally axisymmetric with respective to the
crossings of the intensity distribution of Airy disk.
optical axis, thus the coherent PSF and transfer function can be more
conveniently expressed with the Fourier (inverse) transform in polar
coordinates
the traditional Fourier transform. It is not difficult to prove that the
2𝜋
ℎ(𝑟, 𝜃) = 𝐻 (𝜌, 𝜗)𝑒 𝑗2𝜋𝑟𝜌 cos (𝜃−𝜗)
𝜌𝑑 𝜌𝑑 𝜗 (64) Hankel forward transform has exactly same form as its inverse trans-
∫0 ∫𝜌 form. For an ideal in-focus diffraction-limited optical system, the co-
where (r, 𝜃) are the polar coordinates corresponding to the spatial coor- herent transfer function
( is
) just the pupil function of the objective lens
𝜌
dinates (x, y) 𝐻 (𝜌) = 𝑃 (𝜌) = c𝑖𝑟𝑐 𝑁𝐴∕𝜆
, and the corresponding PSF can be expressed
as
𝑥 = 𝑟 cos 𝜃
( ) ( )
𝑦 = 𝑟 sin 𝜃 (65) 𝜌 𝑁𝐴 𝑁𝐴
ℎ (𝑟 ) = c𝑖𝑟𝑐 𝐽0 (2𝜋𝑟𝜌)2𝜋𝜌𝑑𝜌 = 𝐽1 2𝜋𝑟
∫𝜌 𝑁𝐴∕𝜆 𝑟𝜆 𝜆
Similarly, (𝜌, ϑ) are the polar coordinates corresponding to the spatial ( )2 [ 2𝐽 (𝑟̄) ]
𝑁𝐴 1
frequency coordinates (u, v). When the imaging system is axisymmet- =𝜋 (68)
𝜆 𝑟̄
ric, the coherent transfer function is only related to the variable 𝜌, so
Eq. (64) can be further simplified as
where J1 is the first-order Bessel function of the first kind, as plotted in
ℎ (𝑟 ) = 𝐻 (𝜌)𝐽0 (2𝜋𝑟𝜌)2𝜋𝜌𝑑𝜌 (66) Fig. 24. 𝑟̄ is the normalized version of r, 𝑟̄ = 2𝜋 𝑁𝐴
𝜆
𝑟. Then the intensity
∫𝜌 distribution corresponding to Eq. (68) becomes
where J0 is a zero-order Bessel function of the first kind, which is defined ( )4 [ 2𝐽 (𝑟̄) ]2
𝑁𝐴 1
as 𝐼𝑃 𝑆𝐹 (𝑟) = |ℎ(𝑟)|2 = 𝜋 (69)
𝜆 𝑟̄
2𝜋
1
𝐽 0 (𝑥 ) = 𝑒𝑗𝑥 cos (𝜃−𝜗) 𝑑𝜗
2𝜋 ∫0 Eq. (69) is called intensity PSF, which is usually used to describe the
𝐽 (𝑟̄)
2𝜋 imaging characteristics of an incoherent imaging system. Since 1𝑟̄ is
𝜃 independent integration 1
= 𝑒𝑗𝑥 cos 𝜗 𝑑𝜗 (67) also an Airy function, this intensity distribution is also called “Airy disk”,
2 𝜋 ∫0
as shown in Fig. 25. It can be observed that 83.8% of the incident light
Eq. (66) is so-called Hankel transform, which is the Fourier trans- energy is concentrated in the central bright spot, and the first intensity
form for the axially symmetric function, and is also a special case of zero-crossing appears at the position of 0.𝑁𝐴611𝜆
.
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Similarly, when there is a defocus of distance Δz, the coherent trans- 4.3.1. Contrast transfer function model under the weak object
fer function [Eq. (63)] can be expressed as approximation
√ In Subsection 2.2, we derived TIE in three different ways. However,
𝐻 (𝜌) = 𝑃 (𝜌)𝑒𝑗𝑘Δ𝑧 1−𝜆 𝜌 ,
2 2
all these derivations rely on two restrictive conditions: paraxial approx-
( ) {
𝜌 1 𝜌 ≤ 𝑁𝐴 imation and weak defocusing approximation. TIE is valid if and only
𝑃 (𝜌) = 𝑐 𝑖𝑟𝑐 = 𝜆 (70)
𝑁𝐴∕𝜆 0 else if the two conditions are satisfied simultaneously, and thus, the linear
relationship between the intensity and the quantitative phase can be
Under the paraxial approximation, the angular spectrum diffraction
established. However, these two conditions are sometimes difficult to
term in Eq. (70) can be approximated by the Fresnel angular spectrum
meet in practical experimental conditions. For example, when imaging
transfer function [Eq. (32)]
with a high NA objective, the paraxial approximation is untenable. On
𝐻 (𝜌) = 𝑃 (𝜌)𝑒−𝑗𝜋𝜆Δ𝑧𝜌
2
(71) the other hand, when the defocus distance is too large to meet the weak
Here the z dependence factor has been omitted. It should be noted that defocusing approximation, the phase solved by TIE will also be deviated
in most literature concerning the OTF [250–252], Eq. (71) is often repre- (detailed discussions about this issue can be found in Section 5).
sented in a normalized coordinate system to simplify the relevant anal- It should be noted that there is more than one approach to achieve
ysis and calculation, i.e., the following normalized variables are intro- linearization. The phase contrast transfer function (CTF) method based
duced on weak object approximation is another classical method for determin-
2𝜋 istic phase retrieval [163,221,253–256]. In the following section, we
𝑟̄ = 𝑁𝐴 𝑟 = 𝑘𝑁𝐴 𝑟 (72) will derive the transfer functions of weak objects under coherent illumi-
𝜆
( √ ) nations based on the scalar diffraction theory. When a weak object (the
𝑝𝑎𝑟𝑎𝑥𝑖𝑎𝑙
𝑧̄ = 2𝑘 1 − 1 − 𝑁𝐴2 𝑧 ≈ 𝑘𝑁𝐴2 𝑧 (73) absorption and phase of the object is sufficiently small) is illuminated by
a coherent plane wave, the complex amplitude of the field just behind
𝜌 the object can be simplified as
𝜌̄ = (74)
𝑁𝐴∕𝜆 𝜙(𝐱)≪1
where k denotes the wave number, 𝑃 (𝜌) ̄ is the pupil function of the 𝑈0 (𝐱) = 𝑇 (𝐱) = 𝑎(𝐱) exp [𝑗𝜙(𝐱)] ≈ [𝑎0 + Δ𝑎(𝐱)][1 + 𝑗𝜙(𝐱)] (77)
Δ𝑎(𝐱)≪𝑎0
objective lens, 𝑟̄ is the normalized radical transverse coordinate, 𝑧̄ is ≈ 𝑎0 + Δ𝑎(𝐱) + 𝑗 𝑎0 𝜙(𝐱)
the normalized axial coordinate, and 𝜌̄ is the normalized radial spatial
frequency (more detailed discussions about the normalized coordinate where a0 denotes the direct current (DC) component of the incident
system can be found in Section 8). In such a normalized coordinate sys- plane wave undisturbed by the object. The other two terms represent
tem, the paraxial coherent transfer function of Eq. (67) can be rewritten the diffracted light components. The second term can be further writ-
as ten as Δ𝑎(𝐱) = 𝑎0 𝜂(𝐱), representing the contribution from the absorption
1
variations 𝜂(x); and ja0 𝜙(x) represents the contribution of the phase
̄ 𝑒− 2 𝑗 𝑧̄ 𝜌̄
2
𝐻 (𝜌)
̄ = 𝑃 (𝜌) (75)
component. Taking Fourier transform of Eq. (77), the spectrum of the
Substituting Eq. (75) into Eq. (66), we can get the expression of the transmitted complex amplitude can be expressed as
defocused PSF in the normalized coordinate [ ]
1 𝑈̂ 0 (𝐮) = 𝑎0 𝛿(𝐮)+𝜂( ̂ 𝐮)
̂ 𝐮) + 𝑗 𝜙( (78)
̄ 𝑒− 2 𝑗 𝑧̄ 𝜌̄ 𝐽0 (𝑟̄𝜌)
2
ℎ𝑧̄ (𝑟̄) = 𝑃 (𝜌) ̄ 2𝜋 𝜌𝑑
̄ 𝜌̄ (76)
∫𝜌̄ Substituting Eq. (78) into Eq. (62) and after appropriate simplification,
In fact, the defocused PSF represented by Eq. (76) is also called 3D PSF and the spectral distribution of the intensity image at the propagation
because it describes the 3D complex amplitude distributions in the im- distance z can be obtained as
[ (√ )]
age space formed by an ideal point source. This expression will be dis-
𝐼̂Δ𝑧 (𝐮) = 𝐼0 {𝛿(𝐮) −2 cos 𝑘Δ𝑧 1 − 𝜆2 |𝐮|2 − 1 𝜂(̂ 𝐮)
cussed in detail in Section 8 when we talk about 3D phase imaging.
[ (√ )] (79)
Under coherent illumination, the image formation process is linear
−2 sin 𝑘Δ𝑧 1 − 𝜆2 |𝐮|2 − 1 𝜙( ̂ 𝐮)}
about the complex amplitude. However, due to the effect of the com-
plex conjugation operator [Eq. (60)], the captured intensity is linearly
related neither to the amplitude nor to the phase of the sample, posing a where 𝐼0 = 𝑎20 . Noted that when deriving Eq. (79), the high-order diffrac-
major obstacle for phase retrieval. On the other hand, for an aberration- tion term (the interaction between the intensity component and the
free imaging system with an infinite aperture, (H(u) ≡ 1 and ℎ(𝐱) = 𝛿(𝐱)), phase component is neglected, leaving only their first-order terms) and
𝐼𝑖 (𝐱) = 𝐴2 (𝐱), which means that in the in-focus plane the phase compo- the finite-aperture effect are omitted (assume the objective aperture is
nent of the object cannot be observed at all. Therefore, the basic idea infinite, i.e., P(u) ≡ 1). By further invoking the paraxial approximation,
of the TIE phase retrieval is to generate phase contrast by introducing the angular spectrum diffraction terms in Eq. (79) can be approximated
defocusing into the pupil function, so that the phase information can be by the Fresnel transfer function of Eq. (33), thus Eq. (79) can be further
converted into the intensity information. Then the inverse problem of simplified as [163,221,253]
( ) ( )
phase recovery can be solved by linearizing the quantitative relationship 𝐼̂Δ𝑧 (𝐮) = 𝐼0 {𝛿(𝐮) − 2 cos 𝜋𝜆Δ𝑧|𝐮|2 𝜂( ̂ 𝐮)}
̂ 𝐮) + 2 sin 𝜋𝜆Δ𝑧|𝐮|2 𝜙( (80)
between the intensity and the phase distribution.
It can be seen from the above equation that 𝐼̂Δ𝑧 consists of three
4.3. Image formation model, phase transfer function and linearization components: the DC component I0 𝛿(u) (average intensity or back-
conditions ground intensity), the intensity component associated with the absorp-
( )
tion 2𝐼0 cos 𝜋𝜆Δ𝑧|𝐮|2 𝜂( ̂ 𝐮), and the intensity component associated with
( )
The major obstacle in the phase retrieval problem is the nonlinear the phase 2𝐼0 sin 𝜋𝜆Δ𝑧|𝐮|2 𝜙( ̂ 𝐮). In this case, the intensity now has a lin-
relationship between the intensity (amplitude) and phase of the sample ear relationship with the absorption function a(x) and the phase function
in the image formation process. Although this nonlinear inverse prob- 𝜙(x). The cosine and sine coefficients in front of the absorption function
lem can be solved by the iterative phase retrieval algorithms [59,60,87], and the phase function are usually termed as absorption transfer function
they rely on computationally intensive iterative operations, and there is [ATF, HA (u)] and phase transfer function [PTF, HP (u)], respectively.
no theoretical convergence guarantee. Therefore, deterministic phase ( )
𝐻𝐴 (𝐮) = cos 𝜋𝜆Δ𝑧|𝐮|2 (81)
retrieval methods are more attractive in this regard. The key to deter-
ministic phase retrieval is how to linearize the quantitative relationship
( )
between the intensity and phase information. 𝐻𝑃 (𝐮) = sin 𝜋𝜆Δ𝑧|𝐮|2 (82)
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

The ATF and PTF demonstrate the relative strengths of the In addition, we assume that the absorption is also slowly varying and
information-bearing portion (amplitude or phase) of the image and the satisfies the following first-order Taylor approximation
ever-present background. Although the linearization between the inten- ( )
𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮
𝑎 𝐱± ≈ 𝑎 (𝐱 ) ± ⋅ ∇ 𝑎 (𝐱 ) (86)
sity and the phase has been achieved based on weak object approxima- 2 2
tion, the effects of absorption and phase are still coupled together in Substituting Eq. (85) and Eq. (86) into Eq. (84) yields
the captured intensity images. In order to recover the phase, these two ( ) ( )
𝜆Δ𝑧𝐮 * 𝜆Δ𝑧𝐮
components need to be decoupled. It can be found from Eqs. (81) and 𝑇 𝐱− 𝑇 𝐱+
2 2
(82) that the absorption and phase transfer functions, i.e., cosine and ( ) ( ) { [ ( ) ( )]}
𝜆Δ𝑧𝑢 𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮
sine functions are the even and odd functions of the defocusing distance =𝑎 𝐱− 𝑎 𝐱+ exp 𝑗 𝜙 𝐱 − −𝜙 𝐱+
2 2 2 2
( ) ( ){ [ ( ) ( )]}
Δz, respectively. This indicates that the intensity variation induced by 𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮
≈𝑎 𝐱 − 𝑎 𝐱+ 1+𝑗 𝜙 𝐱− −𝜙 𝐱+
the equal but opposite defocusing is the same for the absorption com- 2 2 2 2
ponents, but is opposite for the phase components. Therefore, it can (87)
be similar to TIE to collect an in-focus image and two defocused images Substituting Eq. (87) into Eq. (84) and then simplifying the equation
with an equal amount and opposite direction as I0 and I ± Δz respectively. using the weak absorption approximation
Thus, similar to the axial intensity derivative estimation in TIE, we can ( ) ( )
𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮
capture two intensity images with equal and opposite defocus. The sub- 𝐼̂Δ𝑧 (𝐮) = 𝑎 𝐱− 𝑎 𝐱+
∫ 2 2
traction of these two images gives a pure phase-contrast image with the { [ ( ) ( )]}
𝜆Δ𝑧𝐮 𝜆Δ𝑧𝐮
effect of the absorption component canceled out 1+𝑗 𝜙 𝐱− −𝜙 𝐱+ exp {−2𝑗𝜋𝐱 ⋅ 𝐮}𝑑𝐱
2 2
𝐼̂Δ𝑧 (𝐮) − 𝐼̂−Δ𝑧 (𝐮) ( ) { }
̂ 𝐮)
=𝐻𝑃 (𝐮)𝜙( (83) ≈ 𝐼̂𝜙=0 (𝐮) + 2 sin 𝜋𝜆Δ𝑧|𝐮|2 ℱ 𝐼0 (𝐱)𝜙(𝐱)
Δ𝑧
4𝐼0
( ) 𝜆Δ𝑧 { [ ]}
+ cos 𝜋𝜆Δ𝑧|𝐮|2 ℱ ∇ ⋅ 𝜙(𝐱)∇𝐼0 (𝐱) (88)
This leads to a linear relationship between the phase and the inten- 2𝜋
sity, and the quantitative phase information can then be reconstructed where
by Fourier space deconvolution with the PTF HP (u). It should also be ( ) ( )
𝜙=0 𝜆Δ𝑧𝐮 * 𝜆Δ𝑧𝐮
noted that since HP (u) tends to oscillate with the increase in the spa- 𝐼̂Δ𝑧 (𝐮 ) = 𝑎 𝐱− 𝑎 𝐱+ exp {−2𝑗𝜋𝐱 ⋅ 𝐮}𝑑𝐱 (89)
∫ 2 2
tial frequency gradually, it contains zero-crossings, especially at higher
represents the Fourier transform of the intensity distribution at a dis-
spatial frequencies. In order to avoid the zero crossings in the pass-
tance Δz when the object phase component is neglected (the object is
band, the defocus distance should be sufficient small Δz → 0 [256,257],
assumed to be purely absorptive). Eq. (88) is called the mixed trans-
and thus the sinusoidal term in the PTF HP (u) can be approximated
( ) fer function proposed by Guigay et al. [259]. Compared with Eq. (80),
as sin 𝜋𝜆Δ𝑧|𝐮|2 ≈ 𝜋𝜆Δ𝑧|𝐮|2 , so that the oscillation can be effectively
these two equations share some similarities. For example, the sine and
avoided. However, it may reduce the phase contrast for lower spatial
cosine terms of ATF and PTF [Eqs. (81) and (82)] are still clearly
frequencies, which compromises the SNR of the phase reconstruction
expressed in Eq. (88). In addition, the mixed model is more compli-
[254,258]. In addition, based on this weak defocusing approximation,
cated, and the phase information is not completely linearized with the
Eq. (83) can be further simplified to TIE with uniform intensity distri-
intensity. Therefore, the phase cannot be directly retrieved based on
bution [Eq. (42)]. We will study this case in detail in Subsection 5.1.
on-step deconvolution based on the PTF, and nonlinear iterative algo-
Another possible solution to zero-crossings is to synthesize and opti-
rithms are needed to get the accurate solution. Unlike the weak ob-
mize the PTF by using more than two images at multiple defocusing
ject approximation, the derivation of the mixed model relies on the
distances [200,256,259,260]. This method allows the response of PTF
slowly varying object approximation, which imposes restrictions on the
to be optimized over a wider range of spatial frequency at the expense
phase/absorption spatial gradients rather than their absolute values.
of additional data acquisition and processing, which will be discussed
Thus, the slowly varying object is a more relaxed and reasonable as-
in Subsection 5.2.
sumption than the weak object approximation [260]. In this case, if
we expand the third term of Eq. (88) by the chain rule (according to
4.3.2. Mixed transfer function model under the slowly varying object [ ] [ ] [ ]
∇2 𝜙(𝐱)𝐼0 (𝐱) = ∇ ⋅ 𝜙(𝐱)∇𝐼0 (𝐱) +∇ ⋅ ∇𝜙(𝐱)𝐼0 (𝐱) ) and combine similar
approximation terms, the following expression can be obtained [259]
As seen above, the TIE or CTF methods rely on different assump- ( ) [ ]
𝜙=0
tions. TIE is valid for short propagation distances and the paraxial ap- 𝐼̂Δ𝑧 (𝐮) = 𝐼̂Δ𝑧 (𝐮) + 2 sin 𝜋𝜆Δ𝑧|𝐮|2 ℱ 𝐼0 (𝐱)𝜙(𝐱)
proximation. On the other hand, the CTF method effectively extends ( ) 𝜆Δ𝑧 { }
+ cos 𝜋𝜆Δ𝑧|𝐮|2 ℱ ∇ ⋅ 𝐼0 (𝐱)∇𝜙(𝐱) (90)
the linearization range from weak defocusing to an arbitrary defocus- 2𝜋
ing distance (no longer requires Δz → 0), but the derivation is based on It is not difficult to prove that under weak defocusing approxima-
the assumption that the object has weak absorption and phase distribu- tion, Eq. (88) can be simplified to the standard TIE [Eq. (10)]. On the
tions. When the object is thick or strongly absorptive, the approximation other hand, if the absorption is assumed to be weak 𝑎(𝐱) = 𝑎0 + Δ𝑎(𝐱),
of Eq. (78) is no longer valid. In order to extend the validity of the ap- Δa(x) ≪ a0 , Eq. (90) can be reduced to the CTF model [Eq. (80)]. This
proximations made in CTF, some researchers attempted to replace the means that the weak phase approximation 𝜙(u) ≪ 1 used in the CTF
weak object approximation with an alternative approximation so-called model [Eq. (80)] can be relaxed to slowly varying phase approximation
“slowly varying object” [221,256,259–262]. First, based on Eq. (20), the | ( ) ( )|
|𝜙 𝐱 − 𝜆Δ𝑧𝐮 − 𝜙 𝐱 + 𝜆Δ𝑧𝐮 | ≪ 1. In particular, the weak phase approx-
Fourier spectrum of the intensity distribution can be represented in the | |
| 2 2 |
differential spatial domain [221] imation requires that the phase value is much smaller than 1, which is
difficult to satisfy for most biological samples. In contrast, the slowly
( ) ( )
𝜆Δ𝑧𝐮 * 𝜆Δ𝑧𝐮
𝐼̂Δ𝑧 (𝐮) = 𝑇 𝐱− 𝑇 𝐱+ exp {−2𝑗𝜋𝐱 ⋅ 𝐮}𝑑𝐱 (84) varying object approximation only requires that the phase variation is
∫ 2 2 much less than 1 within a small spatial neighborhood (± 𝜆Δ𝑧 𝐮
), and thus
2
is more universal than the weak object approximation.
where 𝑇 (𝐱) = 𝑎(𝐱) exp [𝑗𝜙(𝐱)] is the complex transmittance of the object.
In the slowly varying object approximation, the phase change is assumed
4.3.3. Intensity difference model without approximations
to be much smaller than 1 within a small spatial neighborhood (± 𝜆Δ𝑧 𝐮
)
2 In 2015, Sun et al. [206] further proposed an intensity difference
| ( ) ( )| model without approximations. This method is derived based on the an-
|𝜙 𝐱 − 𝜆Δ𝑧𝐮 − 𝜙 𝐱 + 𝜆Δ𝑧𝐮 | ≪ 1 (85)
| | gular spectrum diffraction formula, and invokes no approximations in
| 2 2 |
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Table 4
Comparison of phase retrieval methods and required approximation conditions

Methods Approximation conditions Phase reconstruction algorithms

TIE [141,159,202] Paraxial approximation −𝑘 𝜕𝐼( 𝜕𝑧

𝐱)
= ∇ ⋅ [𝐼(𝐱)∇𝜙(𝐱)]
𝜆2 |u|2 ≪ 1 Fourier solution:
𝜙(𝐱)= − 𝑘ℱ −1 { 4𝜋 2 |u|2 +𝜀 ℱ [ 𝐼(1𝐱) ℱ −1 { 4𝜋 2𝑗2|u𝜋u
𝑗2 𝜋u
|2 +𝜀
ℱ [ 𝜕𝐼(
𝜕𝑧
x)
]}]}
Weak defocusing approximation
𝜀 > 0 is a small constant.
Δz → 0

Weak absorption TIE Paraxial approximation −𝑘 𝜕𝐼(

𝜕𝑧
𝐱)
= 𝐼0 ∇2 𝜙(𝐱)
[158,263] 𝜆2 |u|2 ≪ 1 Fourier solution:
Weak defocusing approximation 𝜙(𝐱)= 𝐼𝑘 ℱ −1 { 4𝜋 2 |u1 |2 +𝜀 ℱ [ 𝜕𝐼(
𝜕𝑧
𝐱)
]}
0
Δz → 0 𝜀 > 0 is a small constant.
Weak absorption approximation
𝑎(𝐱) = 𝑎0 + Δ𝑎(𝐱), Δ𝑎(𝐱) ≪ 𝑎0

CTF [163,221,253] Paraxial approximation (optional) 𝐼̂Δ𝑧 (u)=𝐼0 [𝛿(u) − 2 cos(𝜋𝜆Δ𝑧|u|2 )𝜂( ̂ u) + 2 sin(𝜋𝜆Δ𝑧|u|2 )𝜙( ̂ u)]
𝜆2 |u|2 ≪ 1 Least-square solution:
⎧ ∑𝑁
̂ ∑𝑁
̂ 2 ⎫
⎪ 𝐷(u) 𝑗=1 𝐼Δ𝑧𝑗 (u) sin(𝜋𝜆Δ𝑧𝑗 |u| )−𝐶(u) 𝑗=1 𝐼Δ𝑧𝑗 (u) cos(𝜋𝜆Δ𝑧𝑗 |u| ) ⎪
2

𝜙(𝐱)=ℱ −1 ⎨ 2𝐼0 [𝐶(u)𝐷(u)−𝐸 (u)2 ]+𝜀 ⎬

⎪ ⎪
⎩ ⎭
∑𝑁
𝐶(u) = 𝑗=1 sin (𝜋𝜆Δ𝑧𝑗 |u| )
2 2
∑𝑁
where, 𝐷(u) = 𝑗=1 cos (𝜋𝜆Δ𝑧𝑗 |u| ) 2 2
∑𝑁
𝐸(u) = 𝑗=1 sin(𝜋𝜆Δ𝑧𝑗 |u|2 ) cos(𝜋𝜆Δ𝑧𝑗 |u|2 )
𝜀 > 0 is a small constant.
Solution of of multi-plane symmetrical defocusing:
⎧ ⎫
𝑁
Weak absorption approximation ⎪∑ |sin(𝜋𝜆Δ𝑧𝑗 |u|2 )| [𝐼̂Δ𝑧𝑗 (u)−𝐼̂−Δ𝑧𝑗 (u)] ⎪
2

𝜙(𝐱)=ℱ −1 ⎨ 𝑁 ⎬
∑
𝑎(𝐱) = 𝑎0 + Δ𝑎(𝐱), Δ𝑎(𝐱) ≪ 𝑎0 ⎪𝑗=1 4𝐼0 sin(𝜋𝜆Δ𝑧𝑗 |u|2 ) [|sin(𝜋𝜆Δ𝑧𝑗 |u|2 )| +𝜀] ⎪
2

Weak phase approximation ⎩ 𝑗=1 ⎭

Solution of two[plane symmetrical defocusing: ]
|𝜙(x)| ≪ 1 or sin(𝜋𝜆Δ𝑧|u|2 )[𝐼̂Δ𝑧 (u)−𝐼̂−Δ𝑧 (u)]
Slowly varying phase approximation 𝜙(𝐱)=ℱ −1 4𝐼0 sin2 (𝜋𝜆Δ𝑧|u|2 )+𝜀
or
| ( ) ( )| [ ̂ ]
𝐼 (u)−𝐼̂−Δ𝑧 (u)
|𝜙 𝐱 − 𝜆Δ𝑧𝐮 − 𝜙 𝐱 + 𝜆Δ𝑧𝐮 | ≪ 1 𝜙(𝐱)=ℱ −1 4𝐼Δ𝑧sin(𝜋𝜆Δ𝑧
| | |u|2 )
| 2 2
| 0

𝐼̂Δ𝑧 (u)=𝐼̂Δ𝑧
𝜙=0
(u)+2 sin(𝜋𝜆Δ𝑧u2 )ℱ [𝐼(𝐱)𝜙(𝐱)]
Mixed CTF transfer Paraxial approximation
+ 𝜆Δ𝑧 cos ( 𝜋𝜆Δ𝑧u2 )ℱ {∇ ⋅ [𝜙(𝐱)∇𝐼(𝐱)]}
function [259] 𝜆2 |u|2 ≪ 1 2𝜋
{ ∑Iterative solution:
𝑁 }
𝑗=1 {2 sin(𝜋𝜆Δ𝑧𝑗 u2 )[𝐼̂Δ𝑧𝑗 (u)−𝐼̂Δ𝑧𝜙=0
(u)−Δ𝑧𝑗(𝑛) (u)]}
𝜙(𝑛+1) (𝐱)=ℱ −1 ∑𝑁
𝑗

𝑗=1 4 𝐼 0 sin2
( 𝜋𝜆Δ𝑧𝑗 u 2 )+𝜀

where,
𝜆Δ𝑧𝑗
Δ𝑧(𝑗𝑛) (u) = cos(𝜋𝜆Δ𝑧𝑗 u2 )ℱ {∇ ⋅ [𝜙(𝑛) (𝐱)∇𝐼(𝐱)]}
Slowly(varying ) absorption approximation 2𝜋
𝜀 > 0 is a small constant.
𝑎 𝐱 ± 𝜆Δ𝑧 𝐮
≈ 𝑎(𝐱) ± 𝜆Δ𝑧 𝐮
⋅ ∇ 𝑎 (𝐱 )
2 2 𝜙(n) (x) represents the phase of nth iterations.
Intensity difference None ̂ u) + 𝑅̂ (u)
𝐼̂Δ𝑧 (u) − 𝐼̂−Δ𝑧 (u)= 4𝐴(𝑚𝑛) sin[𝜔(u)]𝜙(
without approximations Iterative
∑
solution:
N
sin[ωj (u)][ÎΔzj (u)−Î−Δzj (u)−R
̂ (n) (u)]
[206] 𝜙(𝑛+1) (x) = arctan{ℱ −1
j=1
∑N (n) }
j=1 {2Am sin[ωj (u)]} +ε
2
√
where, 𝜔𝑗 (u) = 𝑘Δ𝑧𝑗 (1 − 1 − 𝜆2 |u|2 )
𝑅(𝑛) (𝐱) = 𝐴(𝑚𝑛) ℱ −1 {sin[𝜔(u)]ℱ {𝐴(𝑒𝑛) (𝐱)𝑡𝑎𝑛[𝜙(𝑛) (𝐱)]}}
+ℱ −1 {cos[𝜔(u)]ℱ {𝐴(𝑒𝑛) (𝐱)}}ℱ −1 {sin[𝜔(u)]ℱ {𝐴(𝐱) sin[𝜙(𝑛) (𝐱)]}}
−ℱ −1 {sin[𝜔(u)]ℱ {𝐴(𝑒𝑛) (𝐱)}}ℱ −1 {cos[𝜔(u)]ℱ {𝐴(𝐱) sin[𝜙(𝑛) (𝐱)]}}
𝐴(𝑚𝑛) = 𝑀𝑒𝑎𝑛{𝐴(𝐱) cos[𝜙(𝑛) (𝐱)]}, Mean is the mean value.
√
𝐴(𝑒𝑛) (𝐱) = 𝐴(𝐱) cos[𝜙(𝑛) (𝐱)] − 𝐴(𝑚𝑛) with 𝐴(𝐱) = 𝐼(𝐱)
𝜀 > 0 is a small constant.
𝜙(n) (x) represents the phase of nth iterations.

the derivation process, further extend the validity and universality of the
algorithm. From the above discussions, we can summarize the various
linearization methods for phase retrieval. The phase retrieval problem is
an ill-posed inverse problem. In general, we have four different models
to solve the inverse problem, which rely on different assumptions: (1)
TIE; (2) CTF; (3) mixed transfer function model; (4) intensity difference
model without approximations. TIE is only established under the condi-
tions of the paraxial approximation and weak defocusing. The advantage
is that it imposes no restrictions on absorption and phase distributions.
The establishment of the CTF method requires to satisfy the weak object
approximation (or weak absorption and slowly varying phase approxi-
mation) and does not require the paraxial approximation and weak de-
focusing. The mixed model proposed by Guigay et al. [259] relaxes the
limitations of weak defocusing and weak object approximation (but still
needs the paraxial approximation). Furthermore, the method can be re-
duced to TIE when the weak defocus approximation is considered and
is equivalent to the CTF method when incorporating the weak absorp-
tion approximation. Finally, the intensity difference model does not rely
on any approximate conditions. However, the disadvantage is that they
cannot recover the phase deterministically, and a nonlinear iterative so-
lution is needed. In Table 4 and Fig. 26, we summarize the mathemati-
cal models, establishment conditions, the reconstruction algorithms, and
the intrinsic relationship of these methods. Due to the limited space, this
tutorial will not elaborate on their corresponding phase reconstruction
algorithms. The basic idea behind these nonlinear iterative algorithms
is straightforward: they are both based on the principle of “principal
component linearization”. More specifically, the principal components
of the intensity model contributed by the DC term, and the phase term is
initially retained, and more complicated higher-order nonlinear terms
[such as the last cosine term in Eq. (88)] are neglected. In this way,
the initial value can be solved in a linearized manner. Then, the resid-
ual intensity component is considered as a perturbation term and can
be substituted back into the reconstruction equation to further refine
the obtained phase distribution. In general, compared with the iterative
C. Zuo, J. Li and J. Sun et al.
Fig. 26. Mathematical models, establishment conditions, and the intrinsic relationship of different phase retrieval algorithms
phase retrieval method, the convergence of such “principal component
linearization” algorithms is much faster and more stable (usually con-
verges within 3-5 iterations).
4.3.4. Finite aperture effect of imaging systems
When deriving the image formation models and the corresponding
PTFs in previous subsections, the limited aperture effect of the objective
lens is neglected [the objective aperture is assumed infinite, i.e., P(u) ≡ 1
in Eq. (63)]. Obviously, for a practical imaging system, especially for a
microscope, the limited aperture effect of the objective is one of the
key factors that restrict the resolution of phase imaging, so it cannot be
neglected. Fortunately, the effect of the objective NA is very simple for
a coherent imaging system. It can be seen from Eq. (62) that the Fourier
transform of the intensity at the image plane can be expressed as
{ }
| [ ]|2
𝐼̂Δ𝑧 (𝐮) = 𝑈̂ Δ𝑧 (𝐮) ⊗ 𝑈̂ Δ𝑧
∗
(𝐮) = ℱ |ℱ −1 𝑈̂ 0 (𝐮)𝑃 (𝐮)𝐻Δ𝑧 (𝐮) |
| |
where P(u) is the pupil function of the objective lens, and HΔz (u) is
the angular spectrum transfer function [Eq. (63)]. It is not difficult to
find that for an ideal coherent imaging system, the effect of the ob-
jective’s NA is equivalent to a low-pass filter in the frequency domain.
The information with spatial frequencies lower than NA/𝜆 can reach
the image plane, and propagates a distance Δz to form a defocused in-
tensity image [2,4,5]. Thus, when considering the NA of the imaging
system, we can simply assume that the measured object is a low-pass
filtered version of the original ideal object. Then all conclusions drawn
in Subsection 4.3 can be directly applied by incorporating a pupil func-
tion P(u) into the transfer functions. For example, when considering the
limited NA of an imaging system, the ATF and PTF [ Eqs. (81) and (82)]
should be rewritten as
( )
𝐻𝐴 (𝐮) = 𝑃 (𝐮) cos 𝜋𝜆Δ𝑧|𝐮|2 (92)
( )
𝐻𝑃 (𝐮) = 𝑃 (𝐮) sin 𝜋𝜆Δ𝑧|𝐮|2
Optics and Lasers in Engineering 135 (2020) 106187
Their curves are plotted in Fig. 27. Limited by the lens aperture, the
highest spatial frequency of the recovered phase under coherent illumi-
nation is limited to the coherent diffraction limit, i.e., NA/𝜆.
5. Axial intensity derivative estimation
In previous sections, it was shown that the in-focus intensity image
and the axial intensity derivative are required for solving TIE. However,
the axial intensity derivative 𝜕 I/𝜕 z cannot be measured directly, but it
can be estimated from multiple intensity images captured at different
defocus distances. The accuracy of the derivative estimation is critical
for reliable phase retrieval based on TIE. In this section, we will discuss
this issue in detail.
5.1. Two-plane axial intensity derivative estimation
(91) In order to estimate the axial intensity derivative 𝜕 I/𝜕 z, Teague
[141] proposed the classical two-plane finite-difference method in his
original TIE paper. He suggested to capture two intensity images with
slightly defocusing, and the defocus distances of the two images are
equal and opposite relative to the central in-focus image. Then, the
central finite-difference formula is used to estimate the axial intensity
derivative
𝜕𝐼 (𝐱) 𝐼 (𝐱, Δ𝑧) − 𝐼 (𝐱, −Δ𝑧)
≈ (94)
𝜕𝑧 2Δ𝑧
In this section, we explicitly represent the intensity as a function of
the defocus distance Δz. The two-plane-based finite-difference formula
[Eq. (94)] is simple and easy to implement. But it also poses a problem:
how to choose an appropriate defocus distance Δz [141,171,264]. As
shown in the simulation results of Fig. 28, the accuracy of the deriva-
tive estimation increases with the decrease of the defocus distance Δz,
and the spatial resolution of the corresponding reconstructed phase is
also improved. However, when the defocus distance Δz becomes larger,
the accuracy of the differential approximation will decrease, leading to
(93) “phase blurring” effect, i.e., the degradation of the spatial resolution, as
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 27. Transfer functions of a limited aperture imaging system for different defocus distances under coherent illumination. (a) Amplitude transfer function; (b)
Phase transfer function.

Fig. 28. The effect of defocus distances on the TIE phase reconstruction under noise-free condition. (a) Small defocus distance: original intensity image (left); axial
intensity derivative (middle), and recovered phase distribution (right); (b) medium defocus distance; (c) large defocus distance; the square areas with red lines are
magnified for clarity.

shown in Fig. 28(c). The “phase blurring” effect is also referred to as
“non-linearity error”. This is because that the locally linear approxima-
tion is assumed in Eq. (94), and the increase in Δz will inevitably induce
larger errors due to the nonlinear terms of the actual intensity signal.
Therefore, mathematically, the smaller the defocus distance Δz is, the
more accurate the two-plane finite difference can be approximated to
the ideal derivative.
However, real measurements yield data with noise and discretiza-
tion, which makes the problem more complicated. Under noisy condi-
tions, the defocus distance Δz cannot be too small. Otherwise, the in-
tensity derivative estimate will be overwhelmed by noise, as shown in
Fig. 29(a). The cloud-like low-frequency noise will appear in the recon-
structed phase, so the defocus distance Δz has to be large enough in
order to secure an adequate level of SNR, as shown in Fig. 29(b). The
phase blurring effect becomes apparent when the defocus distance Δz
becomes large. Thus a compromise is made where Δz is chosen to bal-
ance the high-order (or non-linearity) error and the noise effect. It is
understandable that the optimal Δz should depend on both the maxi-
mum physically significant frequency of the object and the noise level.
5.1.1. Causes of low-frequency noise and high-frequency phase blurring
From simulation results shown in the previous section, it can be
found that when the intensity images are noisy, the phase reconstructed
by TIE will be contaminated by cloud-like low-frequency artifacts. This
is a notorious problem in TIE, and the causes of the strong sensi-
tivity to low-frequency artifacts will be analyzed in this subsection.
For simplicity, let us consider the simplified TIE under the uniform
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 29. The effect of defocus distance on the TIE phase reconstruction in the presence of noise. (a) Small defocus distance: original intensity image (left), axial
intensity derivative (middle), and recovered phase distribution (right); (b) large defocus distance: original intensity image (left), axial intensity derivative (middle),
and recovered phase distribution (right); the square areas with red lines are magnified for clarity.

intensity [Eq. (42)]

𝜕𝐼(𝐱)
−𝑘 = 𝐼(𝐱)∇2 𝜙(𝐱) (95)
𝜕𝑧
It can be written as ∇−2 ↔ − 21 2 in the Fourier domain
4𝜋 |𝐮|
{ }
̂ 𝐮) 1 1 𝜕𝐼(𝐱)
𝜙( = ℱ
2𝜋𝜆|𝐮|2 𝐼(𝐱) 𝜕𝑧
Δ𝑧→0 1 𝐼̂(𝐮, Δ𝑧) − 𝐼̂(𝐮, −Δ𝑧)
≈ (96)
4𝜋𝜆Δ𝑧|𝐮|2 𝐼̂(𝐮)
̂ 𝐮) is the Fourier transform of the phase to be recovered, and the
where 𝜙(
Fig. 30. Comparison of phase transfer functions of TIE and CTF under different
RHS is the Fourier transform of the normalized axial intensity derivative
defocus distances.
estimated by the two-plane finite difference. Comparing Eq. (96) with
Eq. (83), 𝜋𝜆Δz|u|2 in the denominator of Eq. (96) can be regarded as
the PTF of TIE defocus distance. Comparing Eqs. (97) with (98), it can be found that
𝐻𝑇 𝐼𝐸 (𝐮) = 𝜋𝜆Δ𝑧|𝐮|2 (97) since HTIE (u) is not equal to HCTF (u), the phases reconstructed from the
CTF and TIE may not be identical. In Fig. 30, we compare the curves of
The PTF represents the weights of different spatial frequency compo-
HTIE (u) and HCTF (u) with different defocus distances. It can be seen that
nents of the phase that can be transferred into the intensity via defo-
the two curves almost overlap at low frequencies as Δz approaches to 0.
cusing. The solution to TIE can be interpreted by an inverse filtering
Nevertheless, with the increase in defocus distance Δz, the two functions
process in frequency domain based on the PTF (multiplying the inverse
begin to diverge. The underlying assumption of TIE’s PTF HTIE (u) is the
Laplacian 1/𝜋𝜆Δz|u|2 ). It should be noted that the response of the PTF
phase contrast increases linearly with defocus distance and quadrati-
HTIE (u) decrease to 0 with |u| → 0 quadratically, so the low-frequency
cally with spatial frequencies. This assumption is obviously unreason-
phase components can hardly be transferred into the intensity via de-
able since the phase contrast cannot arbitrarily increase with defocus
focusing. For the limiting case of the zero-frequency component, the
distance or spatial frequencies (violate the energy conservation law).
constant phase does not produce any phase contrast. In other words,
While HCTF (u) is valid for large defocusing, the curve tends to gradu-
since the low-frequency phase information in the captured intensity sig-
ally oscillate with the increase in the spatial frequency and the defocus
nal is weak, it requires a higher gain in the inverse filtering process in
distance Δz. Therefore, the overestimation of the phase contrast at high
frequency domain (the inverse Laplacian has a singularity at the zero-
spatial frequencies is the primary cause of the phase blurring effect if
frequency, and the gain near the zero-frequency tends to infinity accord-
one misuses TIE beyond the small-defocus regime.
ingly). When there is noise, the low-frequency noise components (espe-
cially for the components close to the zero-frequency) will be amplified
by the inverse Laplacian to create cloud-like artifacts superimposed on 5.1.2. Unified framework based on Savitzky-Golay differential filter
the reconstructed phase. In general, the multiple-plane axial intensity derivative estimation
Next, let us turn to the high-frequency phase blurring effect method can improve the high-order error and reduce the noise effect
shown in Fig. 30. For a weak phase object, it can be known from compared to the two-plane method. While these strategies have been
Subsection 4.3.2 that the PTF in the CTF model under weak object ap- shown to work well in many situations, their performance, however,
proximation can be represented as depends heavily on the noise level and the characteristics of the exper-
imental data [197–199]. With a given set of intensity measurements,
𝐻𝐶𝑇 𝐹 (𝐮) = 𝑠𝑖𝑛(𝜋𝜆Δ𝑧|𝐮|2 ) (98) making an appropriate choice of a suitable algorithm is difficult. For the
Note that the establishment of the linearization does not impose any high-order finite-difference with noise-reduction and the least-squares
restrictions on the defocus distance Δz, so Eq. (83) is still valid for a large fitting methods, it is also necessary to determine the fitting order, which
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

is quite similar to the case that how to select the optimal defocus dis-
tance in the two-plane finite-difference method. Therefore, it is neces-
sary to establish a theoretical framework that can more systematically
understand, analyze, compare, and even improve the existing axial in-
tensity derivative estimation methods.
To address this problem, Zuo et al. [205] proposed a unified frame-
work of TIE axial intensity derivative estimation based on the Savitzky-
Golay differential filter. The Savitzky-Golay filter was proposed by Sav-
itzky and Golay [265] in 1964. They proven that fitting a polynomial
to a set of input samples and then evaluating the resulting polyno- Fig. 31. Frequency domain properties of SGDFs with different orders (n=15).
mial at a single point within the approximation interval is equivalent (a) Frequency domain response of the SGDF; (b) frequency domain responses of
to discrete convolution with a fixed impulse response function (i.e., the the low-pass filters implied in SGDFs.
Savitzky-Golay filter). More specifically, the Savitzky-Golay differential
filter (SGDF) is equivalent to the convolutional expression of the least-
squares fitting method proposed by Waller et al. [189]. To evaluate the can be regarded as the combination of an ideal first-order derivative
( )
sth derivative at point t using a polynomial of mth degree on 2𝑛 + 1 data filter (frequency domain response is 𝐻𝑖𝑑𝑒𝑎𝑙 𝑒𝑗𝜔 = 𝑗𝜔) and a low-pass
j 𝜔
filter (frequency domain response is HLP (e )):
points, the coefficients of the SGDF, i.e., the weights of the convolution
kernel can be calculated as [266] ( )
𝑚 ( 𝑗𝜔 ) 𝐻𝑆𝐺 𝑒𝑗𝜔 ( ) ( )
∑ (2𝑘 + 1)(2𝑛)(𝑘) 𝐻𝑆𝐺 𝑒 = 𝑗𝜔 = 𝐻𝑖𝑑𝑒𝑎𝑙 𝑒𝑗𝜔 𝐻𝐿𝑃 𝑒𝑗𝜔 (101)
𝑎𝑖 = 𝑃𝑘𝑛 (𝑖)𝑃𝑘𝑛,1 (0) (99) 𝑗𝜔
𝑘=0 (2𝑛 + 𝑘 + 1)(𝑘+1)
Eq. (101) suggests that SGDF is equivalent to calculating the ideal first-
where (a)(b) is a generalized factorial function (𝑎)(𝑎 − 1)...(𝑎 − 𝑏 + 1) with
order differentiation derivative on a smoothed (denoised) version of the
(𝑎)(0) = 1, and 𝑃𝑘𝑛 (𝑡) is the Gram polynomial, and 𝑃𝑘𝑛,𝑠 (𝑡) is its s-order
data. Higher-order SGDFs can approach the ideal differential filter in the
derivative defined as
( 𝑠 ) higher frequency range in the frequency domain response. However, the
𝑑
𝑃𝑘𝑛,𝑠 (𝑡) = 𝑠
𝑃𝑘𝑛 (𝑥) (100) larger the noise suppression factor, the more sensitive it is to noise. On
𝑑𝑥 𝑥=𝑡
the contrary, although the response of the high-frequency part of the
Simplifying the Eq. (99) by setting the fitting order as 𝑚 =
low-order filter is attenuated, its inherent low-pass filtering has a strong
2𝑛, 1, 𝑚(𝑚 < 2𝑛 + 1) in Eq. (58), Zuo et al. [205] establish the correspon-
resistance to noise. From the frequency response (in log scale) of the
dence between the different finite-difference methods and SGDFs, i.e.,
low-pass filter plotted in Fig. 31, it can be seen the filters have very
least-squares fitting, in particular:
flat frequency responses in their passbands with modest attenuation in
Conclusion 1: The high-order finite-difference method corresponds
their stop band. The frequency response of SGDF is flatter and closer to
to the SGDF with degree 2n.
the ideal derivative filter at low frequencies as the polynomial degree
Conclusion 2: The noise-reduction finite-difference method corre-
increases. The increase of the flatness is coincident with a relatively
sponds to the SGDF with degree 1.
higher cut-off frequency or a wider low-pass range. Conversely, as the
Conclusion 3: The higher order finite-difference with noise-
degree of fitting polynomial decreases, the effect of low-pass filtering
reduction method corresponds to the SGDF with degree m (𝑚 < 2𝑛 + 1).
becomes more evident.
Finally, it should be mentioned that the unequally spaced multi-
In summary, when estimating the axial intensity derivative based
plane methods [265–267] fall in the class of SGDF without exception,
on multi-plane intensity measurements, merely modifying of the finite-
since the Savitzky-Golay filter has also been generalized for unequally or
difference scheme in the spatial domain cannot fully utilize all the in-
non-uniformly spaced data as its offspring. Even the two-plane method
tensity information of multiple planes. Instead, it converts the contra-
can be viewed as a special case of the SGDF when 𝑛 = 1, 𝑚 = 1.
diction in the selection of the defocus distance into the selection of the
The advantage of unifying all of the above derivative estimation
filter degree. A higher-degree SGDF gives a more intact phase (with a
strategies under the framework of SGDF is that it provides great insights
wider range of spatial frequencies can be accurately recovered) than a
into the behaviors, the shortcomings, and the performance of these ex-
lower degree one. However, it results in larger NRR than the one as-
isting intensity derivative estimation algorithms by only analyzing the
sumes a smaller order. In other words, lower-order SGDFs always give
properties of the corresponding SGDFs. SGDF has many good properties
estimates of the low-frequency phase with higher SNR, but they suffer
[268]:
from the problem of poor response for high-frequency phase variations.
1) The convolution operation is simple and is much faster and easier to Therefore, the tradeoff between low-frequency artifacts and the high-
implement than the standard least-squares fitting; order error can still not be fundamentally reconciled by only using a
2) The convolution kernel (weight coefficients) can be easily obtained single degree of SGDF.
using a look-up table generated from the explicit solution, or pre-
calculated using existing routines;
5.1.3. Convergence with the transfer function theory
3) They are optimal differential filters that minimize the NRR subject
TIE is sensitive to noise in the low-frequency range, due to the weak
to moments preservation constraints [267].
transfer of the low-frequency information in the beam propagation. On
Among them, the property (3) is the most important one. From the the other hand, another important aspect is the spatial resolution, which
perspective of probability and statistics: if the signal can be perfectly can be refined if the nonlinear contribution is taken into account. Instead
modeled by an m-order polynomial and the observed data are with in- of optimizing the finite-difference schemes in the spatial domain, many
dependent Gaussian white noise, the mth degree SGDF is an unbiased researchers tried to address the contradiction between low-frequency
estimator of the derivative, and the estimation accuracy can reach the noise and high-order error by optimizing the PTF in the frequency do-
Cramer-Rao lower bound [268]. However, in practice, the signal order main. The optimal estimation of the axial intensity derivative is closely
is unknown, and hence using a nonoptimal degree (i.e., “underfitting” or related to the transfer function theory discussed in Subsection 4.3. The
“over-fitting”) will inevitably affect the accuracy and the effect of noise PTF can be utilized to quantitatively analyze the relationship between
∑
reduction. From the perspective of signal processing, the NRR 𝑛𝑖=−𝑛 𝑎2𝑖 phase contrast and defocus distance, thus providing theoretical guid-
determines the noise resistance ability of the SGDF. Through the fre- ance for optimizing the derivative estimation. There are two options for
quency response decomposition, a SGDF (frequency response HSG (ej𝜔 )) solving this problem:
C. Zuo, J. Li and J. Sun et al.
Fig. 32. The block diagram of the multi-plane TIE phase recovery method based on OFS.
1) Combining TIE with nonlinear phase retrieval algorithms
The nonlinear contributions in the image contrast formation are non-
negligible when large propagation distances are used. The limitation due
to the linearization can be refined by other methods which take into ac-
count the nonlinearity of the phase problem. One of such composite
phase retrieval methods was proposed by Gureyev [269], in order to
improve the phase solution. This method uses the phase map obtained
with TIE as an initial approximation and then is refined by Gerchberg-
Saxton and Fienup method. By using the linear TIE solution as a starting
point, the amount of computations is reduced, and stagnation traps are
avoided. Donnadieu et al. [270] pointed out that the PTF of TIE does
not coincide with the CTF model under weak object approximation at
large defocus distance, so the high-frequency phase information cannot
be reconstructed correctly. He also found that iterative phase retrieval
can effectively alleviate the sensitivity of TIE to low-frequency noise,
and proposed a combined phase recovery scheme similar to Gureyev’s
method [269]. Subsequently, Gureyev et al. [256] unified the first-order
Born approximation and TIE into the same theoretical framework, ex-
panding the linearization range of TIE from the weak defocus to near
Fresnel region. Guigay et al. [259] combined the CTF method with TIE
to derive a mixed transfer function model (discussed in Subsection 4.3).
The model reduces to the standard TIE in the small defocus regime and
is equivalent to the CTF model under the weak object approximation.
Langer et al. [260] compared TIE, CTF, and the mixed model by simu-
lations and experiments. The mixed model seems to be more accurate
and robust to noise than the other two methods, but TIE is most accu-
rate without noise. The mixed model combines the advantages of TIE
for strong absorptive samples and the CTF for large defocusing, effec-
tively extending the valid range of TIE phase retrieval. However, the
mathematical model becomes more complicated and cannot be fully
linearized. Therefore, it is generally necessary to use TIE or the CTF
method to get an initial solution, and then the initial solution is refined
by an iterative procedure to get the final exact solution.
2) Optimum frequency selection based on multi-plane finite-
difference schemes:
The simulation results in Subsection 5.1.1 show that in the tradi-
tional two-plane TIE, when the defocus distance is too small, high-
frequency details can be well preserved, but the low-frequency noise
is strong. In contrast, when the defocus distance is too large, the low-
frequency noise can be suppressed, but high-frequency details vanish.
Although it is difficult to make a good compromise between these two
aspects, it inspires us with a straightforward idea: if the intensity images
Optics and Lasers in Engineering 135 (2020) 106187
at both small and large defocus distances are recorded, we can recon-
struct two different phase images separately. Then, the high-frequency
component of the reconstructed phase with small defocus and the high-
frequency component of the reconstructed phase with large defocus can
be combined to overcome low-frequency noise and high-order error si-
multaneously. Based on this idea, Paganin et al. [264] used a pair of
complementary Gaussian low-pass/high-pass filters to decompose the
phases recovered with both small and large defocusing. Based on the
transfer function analysis, they derived the optimal cut-off frequency of
the complementary filter. Although this method can effectively improve
the noise robustness of TIE, it is only limited to the case of two defo-
cus distances. As an extension of the traditional two-plane TIE, it does
not make full use of all the captured intensity information. Regarding
the multi-plane finite-difference schemes, Zuo et al. [205] analyzed the
PTFs of TIE and CTF, and found that the phase recovered by TIE based
on SGDF can be regarded as a low-pass filtered version of the ideal phase
{ ( ) }
̂ 𝐱) = 𝜙(𝐱) ∗ ℱ −1 𝐻𝐿𝑃 𝑒𝑗𝜔 ||
𝜙( (102)
|𝜔=Δ𝑧𝜋𝜆𝑢2
The frequency response of the low-pass filter HLP is exactly the same as
Eq. (101), suggesting that the frequency characteristic of the retrieved
phase using SGDF is determined by the frequency response of the low-
pass filter that is implicit in the SGDF. Considering the phase informa-
tion lies within a certain frequency bandwidth such that it can be cor-
rectly reconstructed by both a lower degree SGDF and a higher degree
one (here “correct reconstruction” is quantitatively defined in [205] as
to that the spatial frequencies of the phase infomation is below the 0.3dB
cut-off frequency of the low-pass filter HLP (ej𝜔 ) implied in the SGDF with
a specific degree), the lower degree filter is preferred for its lower NRR,
or in another word, higher SNR. Higher-order SGDFs are used if and only
if none of the lower-order filters can correctly reconstruct the phase in-
formation for that given frequency band. Based on this idea, Zuo et al.
[205] proposed the optimal frequency selection (OFS) method. Based on
the unified framework based on SGDF, the optimal cut-off frequencies
for SGDFs with different orders are derived, and the best frequency com-
ponents of the phase images obtained from SGDFs with various degrees
are extracted by a properly designed complementary filter bank. The
extracted spatial frequency components of the phase are finally recom-
bined into an optimal composite phase. Compared with the method of
Paganin et al. [264], OFS is more flexible and can be easily extended to
unequally-spaced and arbitrary number of measurement planes. In addi-
tion, SGDF makes full use of all the intensity measurements for the axial
intensity derivative estimation so that better reconstruction results can
be obtained. However, the disadvantage of this method is that it assumes
C. Zuo, J. Li and J. Sun et al.
that the defocus distances are evenly distributed, so a large amount of
intensity data is still needed. Besides, the selection of the defocus in-
terval was not discussed. To address this problem, Martinez-Carranza
et al. [271] used the noise-reduction finite-difference method of Soto
and Acosta [272] as an example to discuss the optimal defocus distance
selection for multi-plane finite-difference schemes. The selection crite-
rion of the optimal defocus distance can be maximum derivative estima-
tion accuracy [271] or minimum phase reconstruction error [271], and
the latter one proven to provide better phase reconstruction accuracy.
Based on the findings of Falaggis et al. [200], Zhong et al. [273] pro-
posed a Gaussian process (GP) regression algorithm to accurately esti-
mate the axial intensity derivative based on exponentially distributed
defocus intervals. Thereby, the number of required images can be sig-
nificantly reduced without compromising the reconstruction accuracy.
Martinez-Carranza et al. [274] pointed out that the selection of the defo-
cus distance should also be based on the minimum phase reconstruction
error criterion rather than the maximum derivative estimation accuracy
criterion, and further optimized the defocus distances (unequal inter-
vals) based on the OFS principle. Sun et al. [206] further deduced the
nonlinear intensity difference model under non-paraxial and non-weak
object conditions. When the object to be measured has strong absorp-
tion, it firstly adopt the OFS scheme under unequal spacing to obtain
the initial solution. Then the phase is iteratively optimized based on the
nonlinear intensity difference model to obtain a more accurate recon-
struction result.
5.2. Regularization methods for ill-posed inverse problems
No matter which axial intensity derivative estimation methods are
used, the TIE phase retrieval is an inherently ill-posed inverse problem.
Under uniform intensity approximation, TIE becomes a standard Pois-
son equation [Eq. (42)]. The phase retrieval by solving TIE is essentially
a deconvolution based on inverse Laplacian in the frequency domain
[Eq. (96)]. Similarly, in the CTF method discussed in Subsection 4.3,
the defocus-induced phase contrast of a phase object is modeled as the
product of the phase distribution and the PTF. The phase reconstruc-
tion also boils down to a frequency-domain deconvolution based on
the corresponding PTF. Alternatively, the frequency-domain deconvo-
lutions used in these phase retrieval algorithms can be discretized and
represented as an energy minimization problem in the spatial domain
̂ = arg min 1 ‖𝐇𝚽 − 𝐛‖2
𝚽 (103)
𝚽 2 2 tion, we will introduce the basic knowledge about statistics optics and
where 𝐇 ∈ ℝ𝑁×𝑁 represents the matrix representation of the discrete
Laplacian or the corresponding PTF, 𝚽 ∈ ℝ𝑁 represents the phase vec-
tor to be solved, and 𝐛 ∈ ℝ𝑁 represents the measurable quantity (the
normalized axial intensity derivative). In this subsection, we consider
the image contains N pixels and is lexicographically row-stacked into a
single column vector.
As mentioned earlier, due to the presence of noise and the zero-
frequency singularity in the PTF of TIE HTIE (u), the deconvolution is
ill-conditioned: the solution is highly sensitive to small perturbations
(e.g., noise and system misalignment) in the measured data. The ill-
posedness is more pronounced for the CTF method because the PTF
HCTF (u) is strongly oscillatory at higher spatial frequencies, leading to
numerous zero-crossings. The discussions in previous sections focus on
how to minimize the ill-posedness by optimizing the defocus distance or
using multiple intensity measurements. While in mathematics, statistics,
and computer science, an alternative approach is so-called “regulariza-
tion”. The principle of regularization is to convert the ill-posed problem
into a set of well-posed problems by adding prior information to prevent
unrealistic solutions
𝚽̂ = arg min 1 ‖𝐇𝚽 − 𝐛‖2 + 𝜏𝑅(𝚽) (104)
𝚽 2 2
where 𝑅(𝚽) ∈ ℝ𝑁 is the regularization function (also known as the
penalty term), and 𝜏 is the corresponding regularization parameter. Fun-
Optics and Lasers in Engineering 135 (2020) 106187
damentally, the ill-posedness of the deconvolution optimization prob-
lem [Eq. (103)] is due to the fact that the solution space is too wide,
thereby leading to instability. The regularization method, by introduc-
ing additional constraints, defines a compact set enclosing the true so-
lution, then we can search a physically reasonable solution within the
intersection between the original solution space and the defined com-
pact set such that the obtained solution is stable and continuously de-
pendent on the observed data. The well-known and commonly used reg-
ularization models, such as Tikhonov [159,275–277] and total variation
(TV) [275,278,279] were successfully applied to TIE to suppress the low-
frequency artifacts introduced by the ill-posedness of inverse Laplacian.
The most popular Tikhonov regularization approximates the solution
of the inverse problem through a minimization problem by adding a
quadratic regularization term to Eq. (103). It is very simple and easy to
implement. However, the denoising ability is limited and often accom-
panied by the excessive suppression of low-frequency components of the
object. TV regularization is remarkably effective at simultaneously pre-
serving edges whilst smoothing away noise in flat regions, even at low
SNR. It is very suitable for the reconstruction of step-like phase sam-
ples. However, the algorithm is complex in calculation and sensitive to
the selection of regularization parameters, and the reconstructed phase
suffers from the so-called “staircase effect”. On the other hand, prior in-
formation about the sample, such as non-negativity constraints, support
constraints, and the phase-attenuation duality commonly used in X-ray
diffraction imaging [276,280] can also be effectively integrated into the
optimization model. The solution space is further constrained based on
the a priori, and thereby the phase reconstruction quality can be further
improved.
6. Image formation under partially coherent illuminations
No matter completely coherent fields, partially coherent fields, or
completely incoherent fields, the intensity of the optical field is well-
defined and directly accessible. Nevertheless, the definition of “phase” is
only limited to fully coherent fields. Teague’s derivation of TIE was also
based on the assumption of complete coherence [141], i.e., a monochro-
matic coherent beam with a well-defined phase. However, as discussed
in Subsection 4.1, any physically achievable light sources are not strictly
coherent, and in the field of optical microscopy, partially coherent illu-
mination is beneficial to enhance the imaging resolution, improve the
image quality, and suppress the coherent noise. Therefore, in this sec-
coherence theory, and derive the image formation models under par-
tially coherent illuminations. It provides a theoretical basis for the next
section where the generalization of TIE for partially coherent fields is
discussed.
6.1. Correlation function representations of partially coherent fields
The light emitted by most of the light sources in the real world is
generated by different independent radiation oscillators. They produce
fields that vary in time with highly complicated and irregular wave-
forms. Because of diffraction, these waveforms are greatly modified as
the fields propagate. All photodetectors measure the time-averaged in-
tensity over the waveform. This measurement depends on the integra-
tion time of the detector and the waveform of the light at the detector.
Generally, this waveform is not precisely known. Coherence theory is a
mathematical model that is very successful in describing the effects of
this unknown waveform on the observed measurement of time-averaged
intensity. It is based on the electromagnetic wave theory of light as for-
mulated from Maxwell’s equations and uses statistical techniques to an-
alyze the effects due to fluctuations in the waveform of the field in both
time and space. As in scalar diffraction theory, it is much more conve-
nient to treat monochromatic fields than it is to deal with fields that
have complicated time dependencies. Therefore, each of these scalar
components is usually represented at some typical points in space by a
C. Zuo, J. Li and J. Sun et al.
superposition of monochromatic scalar components. Thus the field am-
plitude for a typical monochromatic field U(x, t) with angular frequency
𝜔0 is given by can be expressed as
( ) [ ( )]
𝑈 (𝐱, 𝑡) = 𝑈 (𝐱) exp −𝑗 𝜔0 𝑡 =𝐴(𝐱) exp 𝑗 𝜙 − 𝜔0 𝑡 (105)
The amplitude of the optical field does not change over time, while the
phase changes linearly with time. For a monochromatic deterministic
optical field, the optical oscillation at each spatial point is the same
in time and spatially infinitely extended, therefore, the time-dependent
( )
and fast-fluctuating part exp −𝑗 𝜔0 𝑡 of U(x, t) are usually ignored, it can
be described deterministically by the 2D scalar complex amplitude U(x)
( )
or a “phasor”. In addition, since the exponential term exp −𝑗 𝜔0 𝑡 is the
eigenfunction of the linear time-invariant system, the Fourier transform
of the time variable t for the monochromatic coherent field U(x, t) can
be expressed as
( )
ℱ {𝑈 (𝐱, 𝑡)} = 𝐴(𝐱) exp [𝑗𝜙(𝐱)]𝛿 𝜔 − 𝜔0 (106)
It seems self-evident that there should be only one spectral peak for the
monochromatic optical field. In the following part, we denote the 2D
scalar complex amplitude of the monochromatic coherent field U𝜔 (x)
with angular frequency 𝜔 as the subscript
[ ]
𝑈𝜔 (𝐱) = 𝐴𝜔 (𝐱) exp 𝑗 𝜙𝜔 (𝐱) (107)
However, for practical light sources, from the quantum nature of light
emission, the amplitude and phase of the optical field are inevitably
disturbed under the influence of external temperature, humidity, vibra-
tion, and other factors, together with the statistical fluctuations and in-
homogeneous attenuation of the light-emitting atoms. Therefore, the 2D
complex amplitude function U(x) cannot fully describe the random dis-
turbances at different moments or different spatial positions for partially
coherent fields. For the simplest case, we know that for two beams of
the same frequency U𝜔1 (x) and U𝜔2 (x) meet in space, and whether the
interference fringes can be formed are determined by their interference
[ ]
terms 𝐴𝜔1 (𝐱)𝐴𝜔2 (𝐱) cos 𝜙𝜔2 (𝐱) − 𝜙𝜔1 (𝐱) . If the phase difference between
the two beams cannot keep stable during the observation time due to the
light-emitting mechanism or other factors, the interference term may be
attenuated or even vanish by the random disturbances of the phase in
the time average, thereby directly affecting to the formation of inter-
ference effects. In this case, the complex amplitude representation for
deterministic fields is insufficient. Instead, we need to resort to statisti-
cal approaches to describe the coherence of the field. According to the
theory of stochastic processes, the scalar function U(x, t) is regarded as
a typical member from the “ensemble” that characterizes the statisti-
cal properties at point 𝐱 and time t of the optical field, or so-called a
“realization”. Without loss of generality, the optical field is assumed to
be a stationary and ergodic process, and its statistical properties do not
change over time. Fig. 33 shows a typical “realization” in time (optical
oscillations at a fixed point in space over time) in a partially coherent
field and a typical “realization” in space (optical oscillations at different
locations at a given time point).
From the perspective of temporal frequency (spectrum), the random
disturbance of the optical field U(x, t) can also be regarded as the inco-
Fig. 33. The analytic signal: a typical realization (a) as a function of time for a fixed point in space; (b) as a function of space for one point in time.
Optics and Lasers in Engineering 135 (2020) 106187
herent superimposition of the deterministic monochromatic fields with
different frequencies. So U(x, t) can be written as the following Fourier
transform
𝑈 (𝐱, 𝑡) = 𝑈 (𝐱, 𝜔) exp (−𝑗𝜔𝑡)𝑑𝜔 (108)
∫
where U(x, 𝜔) is the (time-dependence-free) scalar complex amplitude
of the deterministic monochromatic optical field of the frequency 𝜔
[to be distinguished from the monochromatic deterministic field U𝜔 (x),
here U(x, 𝜔) explicitly includes the variable 𝜔 to represent polychro-
matic decomposition of the partially coherent field], and exp (−𝑗𝜔𝑡)
is the corresponding time-dependent part with high-frequency tempo-
ral oscillations. The integral represents the incoherent superposition of
monochromatic optical components of different frequencies. For the de-
scription of the properties of partially coherent fields, U(x, 𝜔) and U(x,
t) are of equal importance (because they are Fourier transform pairs),
which represent a “realization” at the spatial point 𝐱 from the ensemble
of an optical field with an arbitrary state of coherence in the temporal
frequency domain, characterized by the angular frequency 𝜔.
6.1.1. Mutual coherence function and cross-spectral density
As we mentioned in the beginning of this tutorial, the oscillation fre-
quencies of light waves are very high, which is far beyond the temporal
resolution of the current photoelectric imaging device. In fact, the quan-
tity that can be measured is the time-averaged intensity of the optical
field. For example, for a partially coherent field U(x, t), the intensity
captured by photodetector can be represented as
⟨ ⟩
𝐼(𝐱) = |𝑈 (𝐱, 𝑡)|2 = ⟨𝑈 (𝐱1 , 𝑡1 )𝑈 ∗ (𝐱2 , 𝑡2 )⟩||𝐱 =𝐱 =𝐱,𝑡 =𝑡
1 2 1 2
𝑡 =𝑡 1
𝜏=𝑡2 −𝑡1
= ⟨𝑈 (𝐱1 , 𝑡)𝑈 ∗ (𝐱2 , 𝑡+𝜏)⟩||𝐱 (109)
1 =𝐱2 =𝐱,𝜏=0
where the sharp bracket indicates the time average. Similarly, for co-
herent fields, we can create its two copies U1 (x1 , t1 ) and U2 (x2 , t2 ) by
amplitude or wavefront splitting, and then superimpose them in space
to produce interference. The interference pattern captured by the pho-
todetector can be written as [281]
⟨ ⟩
𝐼(𝐱) = |𝑈 (𝐱 , 𝑡 ) + 𝑈 (𝐱 , 𝑡 )|2
| 1 1 2 2 |
𝑡 =𝑡 1
𝜏=𝑡2 −𝑡1 ⟨ ⟩
= |𝑈 (𝐱 , 𝑡) + 𝑈 (𝐱 , 𝑡 + 𝜏)|2
| 1 2 |
⟨ ⟩ ⟨ ⟩
= |𝑈 (𝐱 , 𝑡)|2 + |𝑈 (𝐱 , 𝑡 + 𝜏)|2 +
| 1 | | 2 |
2Re⟨𝑈 (𝐱1 , 𝑡)𝑈 ∗ (𝐱2 , 𝑡 + 𝜏)⟩
= 𝐼(𝐱1 ) + 𝐼(𝐱2 )+2Re⟨𝑈 (𝐱1 , 𝑡)𝑈 ∗ (𝐱2 , 𝑡 + 𝜏)⟩ (110)
Comparing Eq. (109) with Eq. (110), we can find that there is a
common correlation function in both of the intensity distributions. This
function represents the cross-correlation of the light signals at two dif-
ferent points 𝐱𝟏 and 𝐱𝟐 , which is termed as the mutual coherence function
(MCF) [282]. It plays a fundamental role in the coherence theory. For
a stationary and ergodic optical field, MCF is more strictly defined as a
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

correlation function of the optical oscillations at two points x1 and x2

( ) ⟨ ( ) ( )⟩
with a relative time delay 𝜏 [282]. Γ22 (𝜏) = Γ 𝐱2 , 𝐱2 , 𝜏 = 𝑈 𝐱2 , 𝑡 + 𝜏 𝑈 ∗ 𝐱2 , 𝑡 (115)
( ) ⟨ ( ) ( )⟩
Γ12 (𝜏) = Γ 𝐱1 , 𝐱2 , 𝜏 = 𝑈 𝐱1 , 𝑡1 𝑈 ∗ 𝐱2 , 𝑡2 It can be seen that in the SCF, we only consider the influence of
⟨ ( ) ( )⟩
= 𝑈 𝐱1 , 𝑡 𝑈 ∗ 𝐱2 , 𝑡 + 𝜏 (111) relative time delay on the correlation function. Therefore, the SCF is
a direct reflection of the temporal coherence of the optical field. Ob-
where 𝜏 = 𝑡2 − 𝑡1 , the sharp bracket indicates an ensemble average. For
viously, when 𝜏 = 0, Γ11 (0) and Γ22 (0) represent the intensities at the
stationary and ergodic processes, the ensemble average is equivalent
two points x1 and x2 and can be directly measured. It can be seen from
to time average, and thus, the correlation functions in Eqs. (109) and
Eqs. (114) and (115) that the SCF is an autocorrelation function that as-
(110) are identical. Using the definition of MCF, the intensity can also
sumes optical oscillations as a temporal stochastic process, which can be
be simply expressed as 𝐼(𝐱) = Γ(𝐱, 𝐱, 0), and the interference pattern gen-
experimentally measured with use of a Michelson interferometer. The
erated by the superposition of two beams can be expressed as
( ) ( ) ( ) Wiener-Khinchin theorem of the stochastic process tells us that the auto-
𝐼(𝐱) = Γ 𝐱1 , 𝐱1 , 0 + Γ 𝐱2 , 𝐱2 , 0 +2ReΓ 𝐱1 , 𝐱2 , 𝜏 (112) correlation function of a stationary stochastic process is closely related
It should be noted that the MCF is a measurable function that can to its power spectral density (PSD) [285]. For a process that is at least
be characterized through interferometry, as Eq. (112) shows. The co- wide-sense stationary, the autocorrelation function Γx (𝜏) and PSD Sx (𝜈)
herence theory is to describe or characterize the physical characteristics form a Fourier transform pair
of the optical field through its second-order or higher-order statistics
𝑆𝐱 (𝜔) = Γ𝐱 (𝜏) exp (𝑗𝜔𝜏)𝑑𝜏 (116)
that can be experimentally measured. The normalized version of MCF is ∫
also called complex degree of coherence (CDC), which can be defined as
1
[283] Γ𝐱 (𝜏) = 𝑆 (𝜔) exp (−𝑗𝜔𝜏)𝑑𝜔 (117)
2𝜋 ∫ 𝐱
Γ (𝜏) Γ (𝜏)
𝛾12 (𝜏) = √ 12 = √12 (113) where 𝜔 = 2𝜋𝜈 = 2𝜋𝑐 ∕𝜆 is the angular frequency of the optical wave,
Γ11 (0)Γ22 (0) 𝐼1 𝐼2
where 𝜈 is the temporal frequency, c is the speed of light, and 𝜆 is the
MCF and CDC are two very important physical quantities that charac- wavelength. PSD is commonly known as the spectral distribution of the
terize the degree of correlation between optical oscillations at two dif- light source, so the spectral distribution of the source directly deter-
ferent locations in the optical field. It is not difficult to understand from mines the temporal coherence of the optical field. A light source of
Eq. (110) that the physical meaning of CDC represents the contrast of broad spectrum has a low degree of temporal coherence, whereas a
the interferometric fringe pattern formed by the superposition of the light source with narrow linewidth has a high degree of temporal co-
two beams (for example, it can be explained by the Yang’s double slit herence, as illustrated in Fig. 34. The coherence time of a light source
interference experiments of stationary optical fields [284]: x1 and x2 can be increased by using an optical filter to reduce its spectral width.
represent the position of two pinholes, and 𝜏 = 𝑡2 − 𝑡1 represents the The resultant gain of coherence comes at the expense of losing light
time delay between the two beams emitted from the two pinholes at a energy.
given location). The intensity at a given location is proportional to the In order to quantify the temporal coherence more straightforwardly,
beam intensity and the real part of the CDC. It can be proven by using we can define the coherence time and the coherence length based on the
the Schwarz’s inequality that 0 ≤ |𝛾 12 (𝜏)| ≤ 1. It is a properly normalized half-power bandwidth (spectral width) of the source
correlation coefficient, so that 𝛾11 (0) = 𝛾22 (0) = 1. This indicates that the
1
field at a point in space must always be perfectly coherent with itself. 𝜏𝑐 = (118)
Δ𝜈
All other values of 𝛾 12 (𝜏) are generally complex with an amplitude no
more than one. This indicates that the fields at two different points, or at 𝑐 𝜆̄ 2
𝑙𝑐 = 𝑐 𝜏𝑐 = = (119)
the same point after a time delay 𝜏, are generally less than perfectly co- Δ𝜈 Δ𝜆
herent with each other. The magnitude of the complete degree of spatial They are rough measures of the maximum allowable time delay and
coherence (from zero to one) is a measure of the mutual coherence be- path difference between the beam and its replica for forming the in-
tween the fields at the two test points and after a time delay 𝜏. When the terference. Though not as rigorous as the SCF, the coherence time and
value of |𝛾 12 (𝜏)| reaches the maximum, i.e., 1, the optical oscillations coherence length have been widely used due to their simple and in-
of the two points x1 and x2 with a time delay 𝜏 is completely coherent, tuitive physical implications (e.g., the coherence length of a 532 nm
and the fringe contrast reaches the maximum of 1. When the value of continuous laser (0.1 nm spectral width) is 532 nm × 532 nm ÷ 0.1
|𝛾 12 (𝜏)| becomes 0, the optical oscillations of the two points x1 and x2 nm = 2.83 mm). Representative spectral bandwidths for different light
are completely incoherent, i.e., no interference fringes can be observed. sources, and their associated coherence times and coherence lengths are
In the case of 0 < |𝛾 12 (𝜏)| < 1, the optical oscillations of the two points provided in Table 5. Conversely, we can also measure the PSD of the
x1 and x2 are partially coherent. light source by measuring the SCF of the corresponding optical field
We have already discussed in Subsection 4.1 that the coherence of (as illustrated in Fig. 35, the interference patterns generated by the
the optical field is determined by the temporal characteristics (spec- Michelson interferometer with adjustable arm lengths are recorded, the
tral distribution) and spatial characteristics (spatial extension) of the interferometric signal is then converted into the spectrogram by Fourier
light source. Therefore, there exists two types of coherence quantities, transform), which is the basic principle of Fourier transform spectrometer
namely, temporal coherence and spatial coherence, which are included (FTS).
in the concept of MCF. For example, when x1 ≠ x2 , the spatial coherence
of the optical field is included in Γ(x1 , x2 , 𝜏), and when 𝜏 ≠ 0, the tem-
poral coherence of the optical field is included in Γ(x1 , x2 , 𝜏). Although Table 5
MCF completely describes the coherence of the field theoretically, the Spectral widths of typical light sources together with their coherence times and
relationship between spatial coherence and temporal coherence are en- coherence lengths
tangled together. In some cases, we need to analyze the effects of spatial Source Δ𝜈 c (Hz) 𝜏𝑐 =1∕Δ𝜈𝑐 𝑙𝑐 =𝑐 𝜏𝑐
coherence and temporal coherence separately. First, let us consider the ( )
Filtered sunlight 𝜆0 =400 − 800𝑛𝑚 3.74 × 1014 2.67fs 800nm
temporal coherence. In the definition of MCF [Eq. (111)], when the two ( )
Light-emitting diode 𝜆0 =500𝑛𝑚, Δ𝜆0 =50n𝑚 6 × 1013 16.7fs 5𝜇m
points x1 and x2 coincide with each other, MCF degenerates into self Low-pressure sodium lamp 5 × 1011 2ps 600𝜇m
( )
coherence function (SCF) at points x1 and x2 separately [282]. Multimode HeNe laser 𝜆0 =633𝑛𝑚 1.5 × 109 0.67ns 20cm
( )
( ) ⟨ ( ) ( )⟩ Single-mode HeNe laser 𝜆0 =633𝑛𝑚 1 × 106 1𝜇s 300m
Γ11 (𝜏) = Γ 𝐱1 , 𝐱1 , 𝜏 = 𝑈 𝐱1 , 𝑡 + 𝜏 𝑈 ∗ 𝐱1 , 𝑡 (114)
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 34. Two random optical fields with different degrees of temporal coherence. (a) Low temporal coherence: optical ossification, complex degree of temproal
coherence, and power spectral density (from left to right); (b) high temporal coherence: optical ossification, complex degree of temporal coherence, and power
spectral density (from left to right).

The above equation is also referred to as the generalized Wiener-

Khinchin theorem. CSD is a central physical quantity in coherence the-
ory, which is the ensemble-averaged correlation function between a typ-
ical monochromatic component of the field at two different locations 𝐱𝟏
and 𝐱𝟐 . It can also be expressed as
( ) ( ) ⟨ ( ) ( )⟩
𝑊𝜔 𝐱1 , 𝐱2 𝛿 𝜔 − 𝜔0 = 𝑈 𝐱1 , 𝜔0 𝑈 ∗ 𝐱2 , 𝜔0 (121)

where U(x, 𝜔) is the Fourier transform of U(x, t), defined by Eq. (108),
which represents a frequency-domain realization of the ensemble of the
statistic properties of the optical field at the point 𝐱, characterized by the
angular frequency 𝜔. The sharp brackets indicate the ensemble average
for different frequency components. Eq. (121) shows that the compo-
nents of different frequencies in the partially coherent field are uncorre-
lated (unable to form stable interference), so CSD represents the correla-
tion between the optical oscillations of the same frequency component
Fig. 35. The basic principle of Fourier transform spectrometer.
at two different points in space. It quantitatively describes the spatial
coherence of the optical field. A very important property of CSD is that
it is a semi-definite Hermitian matrix. From the basic definition of CSD
[Eq. (120)], the MCF can be expressed as an inverse Fourier transform
of CSD
∞ ( )
Γ12 (𝜏) = 𝑊𝜔 𝐱1 , 𝐱2 exp (−𝑗𝜔𝜏)𝑑𝜔 (122)
∫0

Similar to Eq. (108), Eq. (122) indicates that the MCF of a partially
coherent field results from the superposition of the CSD of different
monochromatic components. Similar to MCF, we can also define the nor-
malized CSD as complex degree of spectral coherence (at frequency 𝜔).
Fig. 36. Michelson stellar interferometer: measuring the spatial coherence of a ( )
( ) 𝑊𝜔 𝐱1 , 𝐱2
quasi-monochromatic field by interferometry to infer the dimension of the light 𝜇𝜔 𝐱 1 , 𝐱 2 = √ ( ) ( )
source. (a) Optical configuration; (b) photograph of a real system. 𝑊𝜔 𝐱1 , 𝐱1 𝑊𝜔 𝐱2 , 𝐱2
( )
𝑊𝜔 𝐱1 , 𝐱2
= √ ( ) ( ) (123)
Next, we turn to the spatial coherence. Similar to the SCF, we can de- 𝑆𝜔 𝐱1 𝑆𝜔 𝐱2
fine the Fourier transform of the MCF as the cross-spectral density (CSD)
[286] where S𝜔 (x) is the PSD, which represents the energy of a particular
monochromatic field component
( ) 1
∞
𝑊𝜔 𝐱1 , 𝐱2 = Γ (𝜏) exp (𝑗𝜔𝜏)𝑑𝜏 (120) 𝑆 𝜔 (𝐱 ) = 𝑊 𝜔 (𝐱 , 𝐱 ) (124)
2𝜋 ∫−∞ 12
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Because different monochromatic field components are uncorrelated means that the quasi-monochromatic field can be approximated to be of
and cannot interfere, the total intensity can be expressed as the sum a single frequency component 𝜔̄ , or equivalently, its temporal coherence
of different monochromatic components is close to that of the monochromatic (fully temporally coherent) field.
∞ Therefore, if we neglect the 𝜔 dependence of W𝜔 (x1 , x2 ), and interpret
𝐼 (𝐱 ) = 𝑆𝜔 (𝐱)𝑑𝜔 (125) it as a quantity which represents the ensemble characteristics of a given
∫−∞
monochromatic field component (𝜔 is not a variable but a constant, i.e.,
From the definition of complex degree of spectral coherence, it is ( )
𝜔̄ ), the MI representation J12 and CSD representation 𝑊𝜔̄ 𝐱1 , 𝐱2 for a
plain to see that this function is a properly normalized correlation co-
quasi-monochromatic field are equivalent in essence
efficient, which is always equal to unity if the two points are brought ( ) ( ) ( )
together, and is always less than or equal to unity as they are separated. 𝑊𝜔̄ 𝐱1 , 𝐱2 ≡ 𝐽 𝐱1 , 𝐱2 ≡ Γ12 𝐱1 , 𝐱2 , 0 (132)
If the magnitude of |𝜇 𝜔 (x1 , x2 )| is unity, it indicates that the monochro-
The temporal coherence is closely related to the spectral distribution
matic field component with angular frequency 𝜔 is perfectly coherent
of the light source. The spectral bandwidth of the light source is a di-
between the two points x1 and x2 . If the magnitude of this function is
rect reflection of the temporal coherence, e.g., the coherence length and
less than unity, it indicates less-than-perfect coherence. If the magnitude
the coherence time. Similarly, the spatial coherence is closely related
is zero, it indicates complete incoherence between the field amplitudes
to the size of the light source (for a quasi-monochromatic incoherent
at the two points. For most partially coherent fields, CSD has signifi-
light source, the CCF of the corresponding far-field diffraction and the
cantly large values only for point separations which keep the two field
intensity of the light source constitute a Fourier transform pair; see dis-
points within the same coherence volume. This function depends only
cussions about the Van Cittert-Zernike theorem in Subsection 6.1.5 for
on the positions of the points and the single angular frequency that the
more details). For an incoherent light source with evenly distributed in-
field components at the two points share. Field components of different
tensity of arbitrarily-shaped area As , we can introduce a rough measure
frequencies are always uncorrelated (and therefore incoherent), even at
of spatial coherence, i.e., coherent area Ac at the distance z from the light
the same point.
source as
6.1.2. Mutual intensity of quasi-monochromatic fields 𝜆̄ 2 𝑧2 𝜆̄ 2
𝐴𝑐 = = (133)
In some earlier literature, in order to more conveniently analyze the 𝐴𝑠 Ω𝑠
spatial coherence, it is generally assumed that the optical field satis- Its physical implication is the maximum allowable area at a plane lo-
fies the so-called “quasi-monochromatic conditions”, i.e., the “narrow-band cated at a distance of z away from the quasi-monochromatic incoher-
condition” (the spectral bandwidth Δ𝜔 of the light source is much smaller ent source within which visible interference fringes can be formed.
than its central frequency 𝜔̄ ) plus “small path difference condition” (the Conversely, we can also measure the spatial coherence of the quasi-
optical path difference within the observation area is much smaller than monochromatic field through interferometry to infer the physical size of
the coherence length of the light source): the light source (as shown in Fig. 35, the interference pattern of the light
Δ𝜔 << 𝜔̄ (126) source is recorded by a Michelson interferometer with an adjustable
arm. When the fringe contrast is reduced from the maximum to 0, the
|𝐱1 − 𝐱2 | << 𝑙𝑐 (127)
| | maximum arm distance is obtained, and thus the angular diameter of
In such cases, Eq. (122) can be simplified as the light source can be calculated), which is the basic principle of the
Michelson stellar interferometer.
∞ ( )
Γ12 (𝜏) = 𝑊𝜔 𝐱1 , 𝐱2 exp (−𝑗𝜔𝜏)𝑑𝜔
∫0
6.1.3. Propagation of mutual coherence function
∞ ( ) For a monochromatic coherent field, the optical field distribution
= exp (−𝑗 𝜔̄ 𝜏) 𝑊𝜔 𝐱1 , 𝐱2 exp [−𝑗 (𝜔 − 𝜔̄ )𝜏]𝑑𝜔
∫0 can be fully described by the complex amplitude distribution, which
exp (−𝑗Δ𝜔𝜏)=1 ∞ ( ) is a function of spatial coordinate. When the complex amplitude is
≈ exp (−𝑗 𝜔̄ 𝜏) 𝑊𝜔 𝐱1 , 𝐱2 𝑑𝜔
∫0 known at one plane, the field distribution at an arbitrary distance Δz
= exp (−𝑗 𝜔̄ 𝜏)Γ12 (0) (128) can be determined based on the scalar diffraction theory discussed in
Subsection 2.2.2. For a partially coherent field, the optical oscillations
The above equation shows that under quasi-monochromatic conditions, at one point in space change irregularly with time, and we need to fo-
whether two points in the field can form interference fully depends on cus on its statistical properties, i.e., investigate the correlation of optical
the spatial coherence. Therefore, the characteristics of the optical field oscillations at two different points in the time-space coordinate system.
at this time can be expressed by the zero-delay MCF Γ12 (0), which is Thus, MCF is the fundamental physical quantity describing the prop-
also called the mutual intensity (MI) J12 at points 𝐱𝟏 and 𝐱𝟐 . erties of the statistical field. The MCF of the optical field may change
( ) ( ) during propagation. In this sense, the coherence of the optical field also
𝐽12 = 𝐽 𝐱1 , 𝐱2 = Γ12 𝐱1 , 𝐱2 , 0
⟨ ( ) ∗( )⟩ propagates along with beam propagation. More specifically, for a par-
= 𝑈 𝐱1 , 𝑡 𝑈 𝐱2 , 𝑡 (129)
tially coherent field propagating in free space, its temporal coherence
Therefore, a quasi-monochromatic optical field can be fully described by does not change because it is only related to the spectral distribution of
a single-frequency harmonic. It is not difficult to find that the MI J12 is the source. However, the spatial coherence changes significantly with
also a Hermitian matrix. Similarly, the complex coherence of the quasi- the propagation of the optical field. Therefore, this section focuses on
monochromatic light is defined as the complex coherence factor (CCF), the transmission characteristics of spatial coherence, i.e., the propaga-
which is the CDC at zero time delay tion of MI/CSD.
In Subsection 2.2.2, we know that the Huygens-Fresnel principle of
Γ12 (0) 𝐽
𝜇12 = = √ 12 = 𝛾12 (0) (130) the wave diffraction is based on the principle of linear superposition:
[Γ11 (0)Γ22 (0)]1∕2 𝐼1 𝐼2 since any complex optical wave field can be interpreted as a collection
Taking Fourier transform of Eq. (128) with respect to the time variable of point sources, it can always be decomposed into a linear combina-
yields tion of simple spherical wavelets. The linear property of the wave equa-
( ) tion allows each spherical wave to propagate independently, and then
𝑊𝜔 𝐱1 , 𝐱2 = 𝐽12 𝛿(𝜔 − 𝜔̄ ) (131)
their contributions on the diffraction plane are superimposed to obtain
The above equation shows that the CSD of the quasi-monochromatic the entire diffraction field. Considering a monochromatic field with a
field contains only one non-zero frequency component located at 𝜔. This complex amplitude distribution U0 (x, t), the complex amplitude of the
C. Zuo, J. Li and J. Sun et al.
optical field after propagating a distance Δz can be expressed as
𝑈Δ𝑧 (𝐱, 𝑡) = 𝑈0 (𝐱′ , 𝑡)ℎΔ𝑧 (𝐱′ , 𝐱)𝑑𝐱′ = 𝑈0 (𝐱, 𝑡) ⊗ ℎΔ𝑧 (𝐱)
∫
where hΔz (x) is the impulse response function for free space propagation
of coherent fields. Under the paraxial approximation, the oblique factor
can be neglected, and the impulse response function becomes spherical
wave represented by Eq. (16) (for quasi-monochromatic fields, 𝜆 should
be replaced by 𝜆). ̄ Based on the scalar diffraction theory and the defini-
tion of MI, we can express the 4D MI function at the plane located at a
distance Δz as
( ) ⟨ ⟩
𝐽Δ𝑧 𝐱1 , 𝐱2 = 𝑈Δ𝑧 (𝐱1 , 𝑡)𝑈Δ𝑧 ∗
(𝐱2 , 𝑡)
⟨ ⟩
= 𝑈0 (𝐱1 , 𝑡)𝑈0∗ (𝐱2 , 𝑡) ℎΔ𝑧 (𝐱′ 1 , 𝐱1 )ℎ∗Δ𝑧 (𝐱′ 2 , 𝐱2 )𝑑 𝐱′ 1 𝑑 𝐱′ 2
∬
( )
= 𝐽0 𝐱1 , 𝐱2 ⊗ ℎΔ𝑧 (𝐱1 , 𝐱2 )
𝐱1 ,𝐱2
where hΔz (x1 , x2 ) is the mutual point spread function for the free space
propagation of coherent fields, which is defined as
ℎΔ𝑧 (𝐱1 , 𝐱2 ) = ℎΔ𝑧 (𝐱1 )ℎ∗Δ𝑧 (𝐱2 )
Therefore, when the 4D MI J0 (x1 , x2 ) at a given plane is known, the
4D MI at a different propagation distance Δz can also be obtained.
Eq. (135) shows that the propagation of MI can also be regarded as a
linear system in 4D space, and the response function of MI for each point
pair in space (x1 , x2 ) is determined by 4D function hΔz (x1 , x2 ). Consid-
ering J0 (x1 , x2 ) as a weight factor, the MI JΔz (x1 , x2 ) at the plane Δz
can be obtained by linear superimposition of all the response functions.
Similar to the angular spectrum diffraction theory discussed in
Subsection 2.2.2, we can also analyze the propagation of partially co-
herent fields in the spatial frequency domain. By performing 4D Fourier
transform of (x1 , x2 ) on both sides of Eq. (135), we can get
( ) ( )
𝐽̂Δ𝑧 𝐮1 , 𝐮2 = 𝐽̂0 𝐮1 , 𝐮2 𝐻Δ𝑧 (𝐮1 , 𝐮2 ) (137)
( ) ( )
where 𝐽̂Δ𝑧 𝐮1 , 𝐮2 , 𝐽̂0 𝐮1 , 𝐮2 and HΔz (u1 , u2 ) are the corresponding 4D
Fourier transforms of JΔz (x1 , x2 ), J0 (x1 , x2 ) and hΔz (x1 , x2 ), respec-
tively. (u1 , u2 ) are the 4D spatial frequency coordinates corresponding
to (x1 , x2 ) in the frequency domain. HΔz (u1 , u2 ) is called the transfer
function of MI for free space propagation, which is defined as
𝐻Δ𝑧 (𝐮1 , 𝐮2 ) = 𝐻Δ𝑧 (𝐮1 )𝐻Δ𝑧
∗
(𝐮2 )
where HΔz (u) is the angular spectrum transfer function for free space
propagation of coherent fields, defined by Eq. (30) [under the paraxial
approximation, its form reduces to Eq. (32)].
6.1.4. Wave equations for the propagation of mutual coherence function
In the previous section, the propagation characteristics of spatial co-
herence was analyzed based on the scalar diffraction theory. However,
it is of some general interest to examine the propagation problem at
a more fundamental level. As we know that the light propagation es-
sentially obeys the wave equation, we can derive a pair of scalar wave
equations governing the propagation of the MCF
1 𝜕2
∇21 Γ12 (𝜏) = Γ12 (𝜏)
𝑐2 𝜕𝜏 2
1 𝜕2
∇22 Γ12 (𝜏) = Γ12 (𝜏)
𝑐2 𝜕𝜏 2
∇2 is the Laplacian operator in 3D space (x, z). Based on the differential
property of Fourier transform, a pair of Helmholtz equations for CSD
propagation are obtained
∇21 𝑊12 + 𝑘2 𝑊12 = 0
∇22 𝑊12 + 𝑘2 𝑊12 = 0
It is easy to verify that MI propagates in accord with the same pair of
Helmholtz equations
∇21 𝐽12 + 𝑘̄ 2 𝐽12 = 0
Optics and Lasers in Engineering 135 (2020) 106187
(134)
(135)
Fig. 37. Gain of coherence by propagation as a result of the spreading of light.
Even if the light is completely incoherent in the source plane, the optical fluctu-
ations at at points x1 and x2 shared a common region, and are therefore partially
correlated.
(136)
∇22 𝐽12 + 𝑘̄ 2 𝐽12 = 0 (144)
where 𝑘̄ = 2𝜋∕𝜆̄ is the average wave number. It can be seen that although
the correlation functions of the partially coherent field are 6D function
defined in two respective 3D spatial coordinates (x1 , z1 ) and (x2 , z2 ),
they also satisfy the Helmholtz equation [Eq. (2)] of the monochromatic
coherent field in each 3D spatial coordinate. In addition, it should be
noted that the Helmholtz equations for the propagation of CSD and MI
share the same form. The only difference is that the general wave num-
ber k used in Eqs. (141) and (142) should be replaced by the average
wave number 𝑘̄ in Eqs. (143) and (144). Because of the similarity be-
tween wave equations of CSD and MI, the properties of MI discussed in
this section can be directly applied to CSD.
6.1.5. The Van Cittert-Zernike theorem
In general, light gains spatial coherence by the mere act of prop-
agation. This is not surprising. Even if the source is completely inco-
herent, the radiation from each point spreads and overlaps with that
(138) from the neighboring points. The light reaching two points in the out-
put plane comes from many points of the input plane, some of which are
common (see Fig. 37). These common contributions create partial cor-
relation between fluctuations at the output points. As one of the most
important theorems of modern optics, the Van Cittert-Zernike theorem
[283,287,288] describes the characteristics of the MI function produced
by a quasi-monochromatic incoherent extended source. As the name
implies, the theorem was first demonstrated in papers by Van Cittert
[287] and Zernike [283,288]. In nearly all optical problems involving
light that does not originate from a laser, the original optical source con-
sists of an extended collection of independent radiators. Such a source
can reasonably be modeled as incoherent in the sense
( ) ( ) ( )
𝐽0 𝐱1 , 𝐱2 = 𝐼0 𝐱1 𝛿 𝐱1 − 𝐱2 (145)
(139) Note that if CSD is used to characterize the source distribution,
( ) ( ) ( )
Eq. (145) should be written as 𝑊0 𝐱1 , 𝐱2 = 𝑆 𝐱1 𝛿 𝐱1 − 𝐱2 . Starting
(140) from Eq. (135), and substituting Eq. (16) into Eq. (145), we can deduce
( ) ( )
𝐽Δ𝑧 𝐱1 , 𝐱2 = 𝐼0 𝐱′ ℎΔ𝑧 (𝐱′ , 𝐱1 )ℎ∗Δ𝑧 (𝐱′ , 𝐱2 )𝑑 𝐱′
∬
( (√ √ ))
exp 𝑗𝑘 Δ𝑧2 + ||𝐱1 − 𝐱′ || − Δ𝑧2 + ||𝐱2 − 𝐱′ ||
2 2
1 ( )
(141) = 𝐼0 𝐱 ′ √ √ 𝑑 𝐱′
𝜆̄ 2 ∬
Δ𝑧2 + ||𝐱1 − 𝐱′ || Δ𝑧2 + ||𝐱2 − 𝐱′ ||
2 2
(142) (146)
The phase factor in the integral can be interpreted as the interfero-
metric superposition of two copies of the spherical wave at points x1 and
(143) x2 emitted from the light source at point x′ in the diffraction plane. The
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

contributions from the interference fields generated by all source points the front and back focal planes of a thin lens, although the physical
are weightedly added based on their respective intensities. Noted that quantities involved are entirely different. We further consider the source
the constant factor is omitted here for simplicity. is quasi-monochromatic and incoherent by substituting Eq. (145) into
Under the paraxial approximation, the spherical wave of the impulse Eq. (150)
response function in the integral can be approximated by the Fresnel im- ( )
( ) ( ) 1 ( ) 2𝜋 ′ ( )
pulse response [Eq. (17)], and the following Van Cittert-Zernike theorem 𝐽𝑓 𝐱1 , 𝐱2 = 𝐽𝑓 𝐱1 − 𝐱2 = 𝐼0 𝐱′ exp − 𝐱 ⋅ 𝐱2 − 𝐱1 𝑑 𝐱′ (151)
[283,287,288] can be obtained 𝜆̄ 2 𝑓 2 ∬ ̄
𝜆𝑓
( ) ( ) Aside from the scaling constants, the MI in the object plane is precisely
𝐽Δ𝑧 𝐱1 , 𝐱2 = 𝐽Δ𝑧 𝐱1 − 𝐱2
( ) the Fourier transform of the source intensity. Note that comparing with
1 ( ′) 2𝜋 ′ ( )
= exp( 𝑗𝜓 ) 𝐼 𝐱 exp − 𝐱 ⋅ 𝐱 − 𝐱 𝑑 𝐱′ (147) Eq. (147), the phase factor exp(j𝜓) disappears. The MI in the object
𝜆̄ 2 Δ𝑧2 ∬ 0 ̄
𝜆Δ𝑧
2 1
plane is a function only of the differences of coordinates in the object
where the phase factor is plane and can easily be found by 2D Fourier transforming the source
[ ( )] intensity distribution.
𝜋 | |2 | |2
exp(𝑗𝜓 ) = exp 𝑗 𝐱
| 1| − 𝐱
| 2| (148)
̄
𝜆Δ𝑧
6.1.6. Coherent mode decomposition
The phase factor can be interpreted as the interferometric superposition Although MCF and CSD can characterize the basic properties of par-
of two copies of the spherical wave at points x1 and x2 emitted from tially coherent fields, including propagation and diffraction. However,
an on-axis point source. The Van Cittert-Zernike theorem, stated mathe- they are both 4D functions, so the relevant analysis and calculation are
matically in Eq. (147), can be expressed in words as follows: aside from quite complicated. Coherent mode decomposition [286,289–291] provides
the factor exp(j𝜓) and scaling constants, the MI in the diffraction plane an effective tool to simplify this problem. It has shown that any par-
is given by a 2D Fourier transform of the intensity distribution across the tially coherent field can be represented as the sum over component fields
source. This relationship can be likened to the relationship between the that are each perfectly self-coherent, but mutually incoherent with each
field across a coherently illuminated aperture and the field observed in other. Thus, the CSD for any field can be represented in the form
the Fraunhofer diffraction pattern of that aperture, although the physi- ( ) ∑ ( ) ( )
cal quantities involved are entirely different. 𝑊𝜔 𝐱1 , 𝐱2 = 𝜆𝑛 (𝜔)𝜓𝑛 𝐱1 , 𝜔 𝜓𝑛* 𝐱2 , 𝜔 (152)
( ) 𝑛
We can also use the intensity in the diffraction plane 𝐼Δ𝑧 𝐱1 =
( ) 1 where 𝜓 n (x, 𝜔) is a complex amplitude called “coherent mode”, and is
𝐼Δ𝑧 𝐱2 = ̄2 2 [i.e., 𝐱1 = 𝐱2 in Eq. (147)] to normalize JΔz (x1 , x2 ) and
𝜆 Δ𝑧 mutually incoherent. 𝜆n is a positive number representing the weight of
rewrite it in terms of CCF
( ) ( )) ′ the corresponding mode. A coherent mode 𝜓 n (x, 𝜔) can be interpreted as
2𝜋 ′ (
( ) ∬ 𝐼0 𝐱′ exp − 𝜆Δ𝑧̄ 𝐱 ⋅ 𝐱2 − 𝐱1 𝑑𝐱 a monochromatic deterministic field of frequency 𝜔 with a well-defined
𝜇Δ𝑧 𝐱1 , 𝐱2 = exp(𝑗𝜓 ) (149) amplitude and phase, and its propagation obeys the Helmholtz equation
∬ 𝐼0 (𝐱′ )𝑑 𝐱′
described by Eq. (2). When 𝐱1 = 𝐱2 , Eq. (152) just represents that the
Note that the phase factor exp(j𝜓) in Eq. (149) does not affect the mod- PSD of the partially coherent field is the incoherent superposition of
ulus of the CCF |𝜇 Δz (x1 , x2 )|, i.e., it does not affect the contrast of the PSDs of different coherent modes
interference fringes produced by two points x1 and x2 in Young’s in- ∑ ∑
terference experiment, so |𝜇 Δz (x1 , x2 )| depends only on the difference 𝑆 𝜔 (𝐱 ) = 𝜆𝑛 (𝜔)𝜓𝑛 (𝐱, 𝜔)𝜓𝑛* (𝐱, 𝜔) = 𝜆𝑛 (𝜔)𝑆𝜔𝑛 (𝐱) (153)
𝑛 𝑛
of coordinates 𝐱2 − 𝐱1 in the diffraction plane. The mathematical state-
ment of the Van Cittert-Zernike theorem is simply a precise statement so 𝜆n can be interpreted as the proportion of the PSD occupied by the
of this relationship between the intensity distribution across the source corresponding coherent mode.
and resulting fringe contrast for given locations of the pinholes. Just as Coherent mode decomposition provides a simple and intuitive way
a point source will create interference fringes of perfect visibility, each to understand the underlying physical mechanism of partially coher-
point on an incoherent source will create a separate fringe of high vis- ent fields. Nevertheless, the coherent mode is derived from the CSD.
ibility. If the source size is too large, these elementary fringe patterns When the CSD is unknown, it is very difficult to obtain the coherent
add with significantly different spatial phases, and the contrast of the mode of a partially coherent field. In order to apply the idea of coher-
overall fringe pattern is reduced. When the propagation distance Δz is ent mode decomposition more conveniently, researchers are not only
large enough, the phase factor exp(j𝜓) disappears, and the far field MI limited to standard the coherent modes extracted from the CSD, but
and the source intensity become a precise 2D Fourier transform pair. directly use some pre-defined non-orthogonal complete basis functions
When studying the properties of a partially coherent imaging system, as “suboptimum” coherent modes to represent the partially coherent
it is often necessary to consider an illumination system with lenses (i.e., field. For example, when studying the multimode output beam of a laser,
condenser), such as the most commonly used Köhler illumination con- Hermite-Gaussian modes [292] or Laguerre-Gaussian modes [293] are
figuration in microscopic imaging systems. In this case, the incoherent often adopted. Hermite-Gaussian modes form an orthogonal basis for
light source (condenser aperture) is placed in the front focal plane of the the solutions to the paraxial wave equation in the Cartesian coordinate
lens (condenser lens) and imaged at infinity. As a consequence, nonuni- system, and the Laguerre-Gaussian modes form an orthogonal basis for
formities of the source brightness distribution are not imaged onto the the solutions to the paraxial wave equation in the cylindrical coordinate
object, and a highly uniform field of illumination is provided. It is not system. The prorogation of these two types of modes can be represented
difficult to prove that in such an optical configuration, the relationship based on the Huygens-Fresnel scalar diffraction theory. Thus, coherent
between the MIs in the front focal plane and the object plane (back focal mode decomposition provides a more convenient way to characterize
plane) of the lens can be represented as the propagation properties of some special partially coherent fields.
{ }
( ) 1 ( ) 2𝜋 [ ′ ( ) ( )]
𝐽𝑓 𝐱1 , 𝐱2 = 𝐽0 𝐱𝟏′ , 𝐱𝟐′ exp 𝐱𝟐 ⋅ 𝐱2 − 𝐱1 − 𝐱𝟏′ ⋅ 𝐱2 − 𝐱1 𝑑 𝐱𝟏′ 𝑑 𝐱𝟐′ 6.1.7. Various models for partially coherent fields
𝜆̄ 2 𝑓 2 ∬ ̄
𝜆𝑓
From Subsection 6.1.5, we know that under the paraxial approxima-
(150)
tion, the CSD in the diffraction plane generated from an incoherent ex-
The above equation shows the MIs in the front and back focal tended light source follows the Van Cittert-Zernike theorem [Eq. (147)].
planes of the thin positive lens form a 4D Fourier transform pair, Specifically, the CSD in the diffraction plane can be generally repre-
( ) ( )| { ( )}|
i.e., 𝐽𝑓 𝐱1 , 𝐱2 = 𝐽̂0 𝐮1 , 𝐮2 | 𝐱1,2 = ℱ 𝐽0 𝐱1 , 𝐱2 | 𝐱1,2 . This re- sented as
|𝐮1,2 =± 𝜆𝑓 |𝐮1,2 =± 𝜆𝑓
̄ ̄
( ) ( ) ( ) ( )
lationship is similar to the relationship between the coherent fields in 𝑊𝜔 𝐱1 , 𝐱2 = 𝜓𝜔 𝐱1 𝜓𝜔* 𝐱2 𝑔 𝐱1 − 𝐱2 (154)
C. Zuo, J. Li and J. Sun et al.
Fig. 38. Physical picture for understanding the physical meaning of the various
models for partially coherent fields [296]. (a) A coherent field; (b) a partially co-
herent field described by the generalized Schell model, where 𝑔̂ (𝛉) is the Fourier
transform of the source distribution (direction factor).
Eq. (154) shares certain similarities with the coherent mode decompo-
sition expressed by Eq. (152). We call 𝜓 𝜔 (x) in Eq. (154) coherent com-
ponent wave. Compared with Eq. (147), its form can be represented as
( )
𝑗 𝜋
𝜓 𝜔 (𝐱 ) = − exp 𝑗 |𝐱|2 (155)
̄
𝜆Δ𝑧 ̄
𝜆Δ𝑧
It can be regarded as the spherical wave at the diffraction plane formed
by an on-axis point source. Different from coherent mode decomposi-
tion that the mode weight can only be a positive constant, the “weight”
( )
𝑔 𝐱1 − 𝐱2 in Eq. (154) is a function of coordinate difference 𝐱1 − 𝐱2 , we
call it direction factor. Compared with Eq. (147), the form of direction
factor g(x) is given by
( )
( ) 2𝜋 ′
𝑔 (𝐱 ) = 𝐼0 𝐱′ exp − 𝐱 ⋅ 𝐱 𝑑 𝐱′ (156)
∬ ̄
𝜆Δ𝑧
In fact, it is just the Fourier transform of the intensity distribution of the
light source, so the direction factor and the CCF are closely related. Gen-
erally speaking, the maximum value of g(x) is located at 𝐱 = 0, and the
decay rate with the two-point distance determines the spatial coherence
of the field.
In Eq. (156), point x′ at the source plane represents the inclined plane
wave component in the corresponding angular spectrum. Although the
coherent component wave in Eq. (154) is fixed, when multiplied with
𝜓 𝜔 (x), its direction is modulated thus it becomes another component
wave of a different direction. According to this idea, we can reinter-
pret the Van Cittert-Zernike theorem of Eq. (147): different point radia-
tors of the incoherent source produce spherical component waves with
different tilted angles. The spherical component waves are weighted
by the direction factor determined by the source intensity distribution
I0 (x′), and then incoherently superposed to create the CSD of the en-
tire field, as illustrated in Fig. 38. When the CSD can be written in the
form of Eq. (154), we consider it satisfying the generalized Schell model
[223,294,295].
If the distance between the source and the diffraction plane is far
enough to satisfy the Fraunhofer approximation, 𝜓 𝜔 (x) can be simplified
as a plane wave. In this case, the generalized Schell model reduces to
the Schell model [294].
( ) √ ( ) ( ) ( ) frequency representation, i.e., the phase-space representation should be
𝑊𝜔 𝐱1 , 𝐱2 = 𝑆𝜔 𝐱1 𝑆𝜔 𝐱2 𝑔 𝐱1 − 𝐱2 (157)
Combined with the definition of Eq. (123), it can be found that in
( )
Schell model, the direction factor 𝑔 𝐱1 − 𝐱2 becomes the CCF 𝜇 𝜔 (x1 ,
x2 ), which only depends on the two-point coordinate difference. It is
easy to verify 𝑔 (𝐱) = 𝑔 ∗ (−𝐱) and 𝑔 (𝟎) = 1.
In the field of laser technology, the Gaussian Schell-model
[286,291] is more commonly used to describe the partially coherent
fields emitted from multimode lasers. GSM is a special case of the Schell
model [Eq. (157)], in which S𝜔 (x) and g(x) both take the forms of Gauss
Optics and Lasers in Engineering 135 (2020) 106187
function
( )
|𝐱|2
𝑆𝜔 (𝐱) = 𝑆𝜔0 exp − (158)
2𝜎𝑠2
( )
|𝐱|2
𝑔𝜔 (𝐱) = exp − (159)
2𝜎𝑔2
where S𝜔0 is a constant representing the central (maximum) PSD value;
𝜎 s and 𝜎 g denote the spatial correlation length and the beam waist width
of the GSM beam, respectively. When 𝜎 s ≫ 𝜎 g , the source is spatially
coherent, and when 𝜎 s ≪ 𝜎 g , the source is almost spatially incoherent.
The ratio between the two factors 𝑞 = 𝜎𝑔 ∕𝜎𝑠 is related to the the well-
known M2 (beam quality factor) [297]
√
2
𝑀 = 1+
2
(160)
𝑞2
When the intensity of the partially coherent field is relatively uniform,
and the spatial coherence of the source is relatively low [|g(x)| decays
rapidly with the increase of the two-point distance], then S𝜔 (x) is a
slowly varying function compared with g(x) so that the following ap-
proximation can be satisfied
√ ( ) ( ) ( )
𝐱 +𝐱2
𝑆𝜔 𝐱1 𝑆𝜔 𝐱2 ≈ 𝑆𝜔 1 (161)
2
Combining Eq. (161) with the Schell’s model [Eq. (157)], the quasi-
homogeneous model is obtained. In the early literature, the sources satis-
fying the quasi-homogeneous model are more intuitively called “slowly
varying homogeneous source”.
( )
( ) 𝐱 +𝐱 2 ( )
𝑊𝜔 𝐱1 , 𝐱2 = 𝑆𝜔 1 𝑔 𝐱1 − 𝐱2 (162)
2
Furthermore, at the end of Subsection 6.1.5, we also consider the case
when the incoherent light source (condenser aperture diaphragm) is
placed at the front focal plane of the lens and imaged at infinity. If the
finite aperture effect of the lens is neglected, the intensity of the illumi-
nation field is perfectly homogeneous. Compared with Eq. (147), it can
be found that 𝜓 𝜔 (x) can be reduced to an plane wave with uniform in-
tensity distribution. In this case, we consider the optical field satisfying
the spatially stationary model (also called statistically stationary model)
( ) ( )
𝑊𝜔 𝐱1 , 𝐱2 = 𝑆𝜔0 𝑔 𝐱1 − 𝐱2 (163)
For a spatially stationary field, it can be interpreted as the incoherent
superposition of plane component waves of different directions with
weighting coefficients determined by the source intensity distribution.
The highly homogeneous illumination field produced by the built-in
Köhler illuminations of most current microscopes can be well described
by the spatially stationary model. The coherence models discussed in
this subsection is summarized in Table 6.
6.2. Phase space representations of partially coherent fields
In addition to correlation functions, another powerful new tool
for understanding and characterizing partially coherent fields is called
“phase-space optics”. As we know, Fourier transform is widely used in
the research of deterministic signals. However, for non-stationary sig-
nals (e.g., partially coherent optical fields), the joint space and spatial-
used. The phase-space optics refers to a representation of optical signals
in an artificial configuration space simultaneously providing informa-
tion about spatial properties of the signal and its angular spectrum. In
1932, Wigner [298] introduced the Wigner distribution function (WDF)
in mechanics that permitted a description of mechanical phenomena in
phase space. Such a Wigner distribution was introduced in optics by
Dolin [299] and Walther [300,301] in the 1960s, to relate partial co-
herence to radiometry. A few years later, the Wigner distribution was
introduced in optics again by Bastiaans [302–304]. He illustrated how
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Table 6
Coherence models of different partially coherent fields

Type of models Cross spectral density Coherent Component

( ) ( ) ( ) ( ) ( )
𝑗 𝜋
Generalized Schell model (Van Cittert-Zernike theorem) 𝑊𝜔 𝐱1 , 𝐱2 = 𝜓𝜔 𝐱1 𝜓𝜔* 𝐱2 𝑔 𝐱1 − 𝐱2 Spherical wave 𝜓𝜔 (𝐱) = − 𝜆𝑧
̄ exp 𝑗 𝜆𝑧
̄ |𝐱 |
2

( ) √ ( ) ( ) (
Schell model (far-field Van Cittert-Zernike theorem) 𝑊 𝜔 𝐱1 , 𝐱2 = 𝑆𝜔 𝐱1 𝑆𝜔 𝐱2 𝑔 𝐱1 − 𝐱2
( ) √ ( ) ( ) (
Gaussian Schell-model 𝑊 𝜔 𝐱1 , 𝐱2 = 𝑆𝜔 𝐱1 𝑆𝜔 𝐱2 𝑔 𝐱1 − 𝐱2
( ) ( ) ( ) 𝑆
𝐱 +𝐱
Quasi-homogeneous model 𝑊 𝜔 𝐱1 , 𝐱2 = 𝑆𝜔 1 2 2 𝑔 𝐱1 − 𝐱2
( ) ( )
Spatially stationary model (Statistically stationary model) 𝑊 𝜔 𝐱1 , 𝐱2 = 𝑆𝜔 0 𝑔 𝐱1 − 𝐱2
) √
Plane wave (nonuniform intensity) 𝜓𝜔 = 𝑆𝜔 (𝐱)
) ( )
Gaussian Beam 𝑆𝜔 (𝐱) = 𝑆𝜔0 exp − 2|𝐱𝜎|2
2
√
Plane wave (smooth intensity) 𝜓𝜔 = 𝑆𝜔 (𝐱)
√
Plane wave (uniform intensity) 𝜓𝜔 = 𝑆𝜔0

space-frequency representations are well-suited for gaining physical in-

sight and developing novel engineering applications for partially coher-
ent imaging. On the other hand, the concept of ambiguity function (AF)
was introduced by Woodward [305] in the signal processing of radar and
sonar measurements in 1953. Twenty years later, Papoulis [306] rein-
troduced this concept into the research of optical imaging system. In
fact, AF and WDF form a Fourier transform pair. Phase space, and in
particular WDF and AF, can be recognized as one common platform for
understanding and applying the physics of more traditional models for
describing electromagnetic signals as they evolve and propagate through
an optical system [307].

6.2.1. Wigner distribution function and ambiguity function

In the classical coherence theory, MCF in the space-time domain and
the CSD in the space-frequency domain are the most commonly used
tools to describe partially coherent fields. However, their inherent bilin-
ear, stochastic, and wave-optical nature often lead to complicated math-
ematics and difficulties in comprehension. The WDF representation can
effectively overcome these shortcomings and provide a particularly con- Fig. 39. The relation between the spatial frequency and the direction of prop-
agation under the paraxial approximation.
venient tool for modeling and analyzing of partially coherent fields due
to its simplicity and intuitiveness. For a partially coherent field U(x, t),
its WDF is defined as the Fourier transform of the corresponding CSD in In geometric optics, an optical ray through one point can be uniquely
a differential space coordinate system determined by its position and direction. Therefore, the WDF can be
( )
𝐱′ 𝐱′ ( ) understood as a more rigorous ray model, defined as “generalized radi-
𝑊𝜔 (𝐱, 𝐮) = 𝑊𝜔 𝐱 + , 𝐱 − exp −𝑗2𝜋𝐮𝐱′ 𝑑 𝐱′ (164)
∫ 2 2 ance” [300,301]. However, different from traditional radiance, general-
ized radiance can be negative. It not only considers that light ray travels
where u is the coordinate in frequency domain corresponding to x. By in straight lines but also accurately describes the wave optical effects of
convention, we denote WDF with the same character W𝜔 , but it should light waves, such as interference and diffraction. More discussion about
be distinguished from CSD by its variables (x, u). There are two points the physical meaning of WDF can be found in Subsection 6.2.6.
need to be noted. On the one hand, similar to the CSD W𝜔 (x1 , x2 ), the When 𝜔 is fixed, the WDF W𝜔 (x, u) can only describe the spatial co-
4D WDF W𝜔 (x, u) only describes the spatial coherence of one single fre- herence of one single frequency component. For the sake of brevity,
quency component 𝜔 in a polychromatic partially coherent field. On the we will omit the subscript 𝜔 and only consider the case of (quasi-)
other hand, the WDF W𝜔 (x, u) is defined in phase space, i.e., it has joint monochromatic field (ignore the temporal coherence). Note that for a
coordinates of both the spatial domain and the spatial frequency domain polychromatic field, the temporal coherence can be simply incorporated
(x, u), which should be distinguished from the CSD W𝜔 (x1 , x2 ) where by the integral over all optical frequencies.
two 2D spatial coordinates are involved. The relationship between the Next, we consider a perfectly coherent (both spatially and tempo-
differential space coordinate system (x, x′) used in WDF and the original ( ) ( ) ( )
rally) field, its CSD can be expressed as 𝑊 𝐱1 , 𝐱2 = 𝑈 𝐱1 𝑈 ∗ 𝐱2 , the
space coordinate system (x1 , x2 ) is defined as corresponding WDF can be expressed as
{ { ( ) ( )
𝐱 +𝐱
𝐱1 = 𝐱 + 𝐱2
′
𝐱 = 12 2 𝐱′ 𝐱′ ( )
𝑜𝑟 𝐱 ′ (165) 𝑊 (𝐱, 𝐮) = 𝑈 𝐱+ 𝑈* 𝐱 − exp −𝑗2𝜋𝐮𝐱′ 𝑑 𝐱′ (167)
𝐱 = 𝐱1 − 𝐱2
′
𝐱2 = 𝐱 − ∫ 2 2
2

Note that this is a unitary transformation and the corresponding Jaco- Similarly, we can define the AF
( )
bian is equal to unity. It can be seen that the WDF arises “midway” be- ( ) 𝐱′ 𝐱′ ( )
𝐴 𝐮′ , 𝐱 ′ = 𝑊 𝐱 + ,𝐱 − exp −𝑗2𝜋𝐮′ 𝐱 𝑑𝐱 (168)
tween the CSD W𝜔 (x1 , x2 ). The physical meaning of WDF can be strictly ∫ 2 2
understood as the probability density distribution of photon position and
It is not difficult to prove that AF and WDF form a Fourier transform
momentum. In Subsection 2.2.2, we know that the spatial frequencies
pair:
of the scalar coherent field correspond to the plane wave components
(angular spectrum) propagating in different directions, so we can (ap- ( ) [ ( )]
𝐴 𝐮′ , 𝐱 ′ = 𝑊 (𝐱, 𝐮) exp −𝑗2𝜋 𝐮′ 𝐱 − 𝐮𝐱′ 𝑑𝐱𝑑 𝐮
proximately) interpret 𝐮 of the WDF W𝜔 (x, u) as the direction of light ∬
( )
rays at point 𝐱. Considering the 1D simplified case shown in Fig. 39, the = ℱ {𝑊 (𝐱, 𝐮)} 𝐮′ , 𝐱′ (169)
spatial frequency u of WDF and the ray direction (the angle 𝜃 between
the ray and optical axis) has the following relationship Therefore, as two central physical quantities in phase space optics, the
WDF and the AF representations are equivalent in essence. In the next
cos 𝛼 sin 𝜃 𝑝𝑎𝑟𝑎𝑥𝑖𝑎𝑙 𝜃
𝑢= = ≈ (166) subsection, we will focus on the properties of WDF.
𝜆 𝜆 𝜆
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

6.2.2. Properties of Wigner distribution function 1) Fresnel propagation

The popularity of WDF for characterizing partially coherent fields
roots from its unique properties. In this subsection, we give a brief in- Under the paraxial approximation, the Fresnel propagation property
troduction to the most commonly used properties of WDF. of WDF can be derived from the Fresnel propagation formula of coherent
fields [Eq. (18)] and partially coherent fields [Eq. (135)]
1) Realness
𝑊Δ𝑧 (𝐱, 𝐮) = 𝑊0 (𝐱 − 𝜆Δ𝑧𝐮, 𝐮) (177)
W(x, u) is always a real function
which is an x-shear of the input WDF, and the amount of shear is linearly
𝑊 (𝐱, 𝐮) ∈ ℝ, ∀𝐱, ∀𝐮 (170) proportional to the propagation distance.
This property can be derived from the positive semidefiniteness and 2) Chirp modulation (lens)
Hermitian symmetry of CSD. Note that, W(x, u) is not necessarily non-
negative; this prohibits a direct interpretation of the WDF as an energy When the optical field passes through
( a)lens or a quadratic phase
𝜋
density function (or radiance function). mask, a quadratic phase factor exp 𝑗 𝜆𝑓 |𝐱|2 is induced, where f is the
focal length representing the curvature of the phase factor. The WDF of
2) Spatial marginal property
the output field can be expressed as
( )
𝐱
𝑊𝜔 (𝐱, 𝐮)𝑑𝐮 = 𝑊𝜔 (𝐱, 𝐱) = 𝑆𝜔 (𝐱) (171) 𝑊𝑓 (𝐱, 𝐮) = 𝑊0 𝐱, 𝐮 + (178)
∫ 𝜆𝑓

The equation shows that the integral of WDF over the frequency vari-
able represents the PSD at optical frequency 𝜔 and position x. When the
signal is quasi-monochromatic, the spatial marginal of WDF is just the
intensity. More strictly speaking, for polychromatic fields, the intensity
is the total energy obtained by integrating the PSD over all wavelengths
𝑊𝜔 (𝐱, 𝐮)𝑑 𝐮𝑑 𝜔 = 𝑆𝜔 (𝐱)𝑑𝜔 = 𝐼 (𝐱) (172)
∬ ∫
3) Spatial frequency marginal property
which is a u-shear of the input WDF, and the amount of shear is inversely
proportional to f.
3) Fourier transform (Fraunhofer diffraction)
The WDF corresponding to the Fourier transform 𝑈̂ (𝐮) of the origi-
nal signal U(x) can be realized by interchanging the space and spatial
frequency coordinates of the original WDF.
𝑊𝑈̂ (𝐱, 𝐮) = 𝑊𝑈 (−𝐮, 𝐱) (179)
4) Fractional Fourier transform

𝑊𝜔 (𝐱, 𝐮)𝑑𝐱 = 𝐺𝜔 (𝐮) (173) Fractional Fourier transform is an extension of traditional Fourier
∫
transform to fractional order [309–312], which is generally defined as
G𝜔 (u) is the Fourier transform of the spectral density, called direc- a function of rotation angle 𝜃
tional power spectrum. For a monochromatic coherent field U(x), G(u) { ( )}
is the power spectrum in the spatial frequency domain, i.e., |U(u)|2 . 𝑓𝜃 (𝐱) = ℱ𝜃 𝑓 𝐱𝑖𝑛
( ) ( )
⎡ 1 ⎤ ⎡ |𝐱𝑖𝑛 |2 + |𝐱|2 cos 𝜃 − 2𝐱𝑖𝑛 ⋅ 𝐱 ⎤
4) Convolution property ⎢ exp 𝑗 2 𝜃 ⎥ ( ) ⎢ | | ⎥
= ⎢ √ ⎥ 𝑓 𝐱𝑖𝑛 exp ⎢𝑗𝜋 ⎥𝑑 𝐱𝑖𝑛 (180)
⎢ 𝑗 sin 𝜃 ⎥ ∫ ⎢ sin 𝜃 ⎥
When 𝑈 (𝐱) = 𝑈1 (𝐱)𝑈2 (𝐱), then ⎣ ⎦ ⎣ ⎦
𝑊 (𝐱, 𝐮) = 𝑊1 (𝐱, 𝐮)⊗ 𝑊2 (𝐱, 𝐮) (174) where 𝜃 is the rotation angle and 𝜃 ≠ n𝜋. The fractional order of
𝐮
the transform is defined as 𝜃/(𝜋/2). In particular, when 𝜃 = 0, 𝑓 (𝐱) =
where ⊗ indicates convolution over 𝐮, W1 (x, u) and W2 (x, u) are the
𝐮 ℱ0 {𝑓 (𝐱)} is the original signal; when 𝜃 = 𝜋, 𝑓 (−𝐱) = ℱ𝜋 {𝑓 (𝐱)} is the
WDFs of U1 (x) and U2 (x), respectively. mirror image of the original signal; when 𝜃 = 𝜋∕2, 𝑓̂(𝐮) = ℱ𝜋 {𝑓 (𝐱)}
When 𝑈 (𝐱) = 𝑈1 (𝐱) ⊗ 𝑈2 (𝐱), then is the traditional (1st order) Fourier transform of the original signal;
2
𝐱 { }
𝑊 (𝐱, 𝐮) = 𝑊1 (𝐱, 𝐮)⊗ 𝑊2 (𝐱, 𝐮) (175) when 𝜃 = −𝜋∕2, 𝑓 (𝐱) = ℱ− 𝜋 𝑓̂(𝐮) is the traditional inverse (-1st or-
2
𝐱 der) Fourier transform of the original signal. In optics, the fractional
⊗ indicates the convolution over 𝐱. Fourier transform can be realized by single lens or double lens setups
𝐱
[respectively referred to as type I (RQR setup) and type II (QRQ setup)]
5) Instantaneous frequency [310,311], as shown in Fig. 40. By changing the focal length or spac-
ing of the lens [for type I setup 𝑅 = tan (𝜃∕2); 𝑄 = sin 𝜃, for type II setup
For monochromatic coherent fields 𝑈 (𝐱) = 𝑎(𝐱)𝑒𝑗𝜙(𝐱) , the instanta-
neous frequency of U(x) is associated with the phase gradient ∇𝜙(x)
(the concept of “instantaneous frequency” for 1D signal [308] is adopted
here to represent the phase gradient, see Subsection 6.2.2 for detailed
discussions), which has the following relationship with the WDF
∫ 𝐮𝑊 (𝐱, 𝐮)𝑑𝐮 1
= ∇𝜙(𝐱). (176)
∫ 𝑊 (𝐱, 𝐮)𝑑𝐮 2𝜋
For the proof of Eq. (176), readers can refer to [307].

6.2.3. Optical transformation of Wigner distribution function

The concept of WDF is very close to the optical ray in geometric op-
tics. Although it cannot be fully equivalent to energy density function
in a strict sense, its propagation and optical transformation properties
strictly follow the ray model, making it a bridge between geometric op-
tics (radiometry) and wave optics. In this subsection, we outline the
optical transformation properties of WDF.
Fig. 40. Fractional Fourier transform. (a) Setup (Type I) for performing a frac-
tional Fourier transform. Parameters R and Q determine the degree P and the
angle 𝜙 = 𝑃 𝜋∕2. The singals are 2D; the lens is spherical; (b) setup (Type II) for
performing a fractional Fourier transform.
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

where the ABCD-matrix is symplectic and the determinant equals unity.

It can be seen that the five fundamental optical transformations de-
scribed above can be regarded as special cases of the first-order optical
system, and their corresponding ABCD-matrices are
[ ] [ ] [ ] [ ] [ ]
1 −𝜆𝑧 1 0 0 −1 cos 𝜃 − sin 𝜃 𝑀 0
, , , , (185)
0 1 (𝜆𝑓 )−1 1 1 0 sin 𝜃 cos 𝜃 0 1∕𝑀

The typical optical transformations in phase space are summarized in

Fig. 42.

6.2.4. Wigner distribution representations of typical optical signals

We shall illustrate the concept of WDF by some typical examples
from Fourier optics. For monochromatic deterministic fields, i.e., it
can be completely represented by a given complex amplitude function
𝑈 (𝐱) = 𝑎(𝐱)𝑒𝑗𝜙(𝐱) . The WDFs of five typical coherent signals are shown
in Fig. 43.

Fig. 41. Beam amplification (compression). Output plane is located at 𝑧2 = 𝑧1 + 1) Point source
(1 + 𝑓 ∕𝐹 )𝑓 . The output-plane wave amplitude is 𝑈𝑧2 (𝐱) = (−1∕𝑀 )𝑈0 (−𝐱∕𝑀),
where 𝑀 = 𝑓 ∕𝐹 . A monochromatic point source located at x0 can be expressed by the
( )
impulse signal 𝑈 (𝐱) = 𝛿 𝐱 − 𝐱0 . Its WDF takes the form
( )
𝑊 (𝐱, 𝐮) = 𝛿 𝐱 − 𝐱0 (186)
𝑄 = tan (𝜃∕2); 𝑅 = sin 𝜃], fractional Fourier transform pairs of different
orders can be obtained at the input plane and the output plane. This is a 2D section of the 4D phase space in 𝐱-plane. For the simplified
It can be proven that the WDF corresponding to the fractional Fourier 1D case [see Fig. 41(a)], it corresponds to a line perpendicular to the
transform 𝑈̂ 𝜃 (𝐮) of the original signal U(x) can be expressed as phase- 𝐱-axis and at a distance x0 from the 𝐮-axis in phase space.
space rotation of the original WD
2) Plane wave
𝑊𝑈̂ (𝐱, 𝐮) = 𝑊𝑈 (𝐱 cos 𝜃 − 𝐮 sin 𝜃, 𝐮 cos 𝜃 + 𝐱 sin 𝜃) (181)
𝜃
A plane wave with a spatial frequency u0 can be described by 𝑈 (𝐱) =
Similarly, we can take a projection of 𝑊𝑈̂ (𝐱, 𝐮) over 𝐮, the obtained exp(𝑗2𝜋𝐮0 𝐱). Its WDF takes the form
𝜃
intensity ∫ 𝑊𝑈̂ (𝐱, 𝐮)𝑑𝐮 is actually the Radon transform of the orig- ( )
𝜃 𝑊 (𝐱, 𝐮) = 𝛿 𝐮 − 𝐮0 (187)
inal WDF WU at angle 𝜃 (see Subsection 6.2.5 for details) [313].
∫ 𝑊𝑈̂ (𝐱, 𝐮)𝑑𝐮 is also known as the generalized marginal or generalized This is a 2D section of the 4D phase space in 𝐮-plane. For the 1D case
𝜃
projection of WDF, which is proven to be nonnegative. If we change the [see Fig. 43(b)], it corresponds to a line perpendicular to the 𝐮-axis and
fractional order (rotation angle 𝜃) of the Fourier transform and collect at a distance u0 from the 𝐱-axis in phase space.
the corresponding intensity distributions, we can reconstruct WDF based
on the principle similar to computerized tomography (CT). Such a coher- 3) Spherical wave
ence measurement technique, which reconstructs WDF by phase-space A spherical wave under the paraxial approximation can be de-
rotation and projection, is called phase-space tomography [314–317], scribed by a quadratic-phase signal 𝑈 (𝐱) = exp(𝑗2𝜋𝑎𝐱2 ). Its WDF takes
(relevant contents will be discussed in Subsection 6.2.6). the form
5) Beam amplifier (compressor) 𝑊 (𝐱, 𝐮) = 𝛿(𝐮 − 𝑎𝐱) (188)
If the optical signal passes through a 4f optical system with two lens For the 1D case, it corresponds to a straight line across the origin of
of different focal lengths (Fig. 41), the size of the beam will be magnified phase space [see Fig. 43(c)].
(or compressed) M times (M is the focal length ratio between two lenses),
then the WDF of the output field can be represented by the scaling of 4) Slow varying wave
the original WDF For a smooth phase signal in the spatial domain, 𝑈 (𝐱) = exp [𝑗𝜙(𝐱)]
( )
𝐮 where 𝜙(x) is a smooth function of 𝐱, its WDF takes the form [321]
𝑊𝑀 (𝐱, 𝐮) = 𝑊0 𝑀𝐱, (182)
𝑀 ( )
1
𝑊 (𝐱, 𝐮) ≈ 𝛿 𝐮 − ∇𝜙 (189)
6) First order optical system 2𝜋
For more general slowly varying field 𝑈 (𝐱) = 𝑎(𝐱) exp [𝑗𝜙(𝐱)] (the ampli-
The first-order optical system is one of the most common optical systems
tude a(x) is a nonuniform but smooth function of 𝐱). Its WDF takes the
[318]. Similar to the matrix transformations in geometric optics, the ge-
form [208]
ometric optical characteristics of the axisymmetric paraxial optical sys-
( )
tem can be described by a 2 × 2 ABCD-matrix [319]. Similarly, in wave 1
𝑊 (𝐱, 𝐮) ≈ 𝐼 (𝐱)𝛿 𝐮 − ∇𝜙 (190)
optics, the transformation of a first-order optical system on the input 2𝜋
field can also be described by Collins diffraction integral transformation More details about Eq. (190) will be introduced in Subsection 7.2.4. For
equation based on ABCD-matrix [320]. Similar to the ABCD-matrix in the 1D case, it corresponds to a curve in phase space [see Fig. 43(d)].
matrix optics, the WDF of the original signal U after passing through a The proof of Eq. (190) can be found in [208], and the required “slow
first-order optical system can be expressed as varying” condition is strictly defined as
( ) ( )
𝑊𝑈1𝑠𝑡 (𝐱, 𝐮) = 𝑊𝑈 (𝐴𝐱 + 𝐵𝐮, 𝐶𝐱 + 𝐷𝐮) (183) 𝐱′ 𝐱′
𝜙 𝐱+ −𝜙 𝐱− ≈ 𝐱′ ⋅ ∇𝜙(𝐱) (191)
2 2
In phase space, the transformation between input and output coordi-
( ) ( )
nates is 𝐱′ 𝐱′
[ ′] [ ][ ] 𝑎 𝐱+ 𝑎 𝐱− ≈ 𝑎2 (𝐱) = 𝐼 (𝐱) (192)
𝐱 𝐴 𝐵 𝐱 2 2
= (184)
𝐮 ′ 𝐶 𝐷 𝐮
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 42. Typical optical transformations in phase space. (a) Fresnel propagation; (b) Chirp modulation (lens); (c) Fourier transform; (d) fractional Fourier transform;
(e) magnifier.

Fig. 43. WDFs of typical coherent signals. (a) Point source; (b) plane wave; (c) spherical wave; (d) slow-varying wave; (e) Gaussian beam.

5) Gaussian beam or WDF. Here, we introduce the WDFs of three typical partially coherent
{ fields.
( )2 }
The WDF of the Gaussian signal 𝑈 (𝐱) = exp − 𝜎𝜋2 𝐱 − 𝐱0 is
1) Spatially incoherent field
{ ( )}
𝜋 ( )2 𝜎 2 2 For spatially incoherent fields, the CSD can be expressed as
𝑊 (𝐱, 𝐮) = exp − 𝐱 − 𝐱0 + 𝐮 (193)
𝜎2 2𝜋 [Eq. (145)]
( ) ( ) ( )
which is still a Gaussian signal in both 𝐱 and 𝐮. For the 1D case, it cor- 𝑊𝜔 𝐱1 , 𝐱2 = 𝑆𝜔 𝐱1 𝛿 𝐱1 − 𝐱2 (194)
responds to a 2D Gaussian function in phase space [see Fig. 43(e)].
For quasi-monochromatic spatially incoherent fields
For partially coherent fields, it cannot be described by the complex
( ) ( ) ( )
amplitude. Its statistical characteristics need to be represented by CSD 𝑊 𝐱1 , 𝐱2 = 𝐼 𝐱1 𝛿 𝐱1 − 𝐱2 (195)
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

where the PSD S𝜔 (x) or intensity I(x) is a nonnegative function. In the equation and combining the properties of WDF under Liouville approxi-
central differential space coordinate defined by Eq. (165), the CSD be- mation (geometric optics approximation), we can deduce that the prop-
comes agation of WDF follows the following Liouville transport equation [322]
( ) √
𝐱′ 𝐱′ ( )
𝑊𝜔 𝐱 + , 𝐱 − = 𝑆𝜔 (𝐱)𝛿 𝐱′ (196) 𝑘2 − 4𝜋 2 |𝐮|2 𝜕𝑊 (𝐱, 𝐮)
2 2 = −𝜆𝐮∇𝐱 ⋅ 𝑊 (𝐱, 𝐮) (204)
𝑘 𝜕𝑧
Its corresponding WDF is ( )
where ∇𝐱 = 𝜕 ∕𝜕 𝐱 = 𝜕𝑥 , 𝜕𝑦 . This equation can be solved analytically,
𝑊𝜔 (𝐱, 𝐮) = 𝑆𝜔 (𝐱) (197) and the solution takes the form
which is a function only of the space variable 𝐱 and that it does not ⎛ ⎞
depend on 𝐮. This suggests that the power density of spatially incoherent ⎜ 2𝜋𝐮 ⎟
𝑊𝑧 (𝐱, 𝐮) = 𝑊0 ⎜𝐱 − √ 𝑧, 𝐮⎟ (205)
field is independent with the spatial frequency (ray direction). ⎜ 𝑘2 + 4𝜋 2 |𝐱|2 ⎟
⎝ ⎠
2) Spatially stationary field Eq. (205) describes the general propagation law of WDF in free space.
[ ]𝑇
The CSD of quasi-monochromatic fields satisfying the spatially sta- Based on WDF, we can construct a 3D vector field 𝑗𝐫 = 𝑗𝐱 , 𝑗𝑧 , which
tionary model can be expressed as [Eq. (163)] is known as the geometrical vector flux [303,323]
( ) ( )
𝑊𝜔 𝐱1 , 𝐱2 = 𝑆0 𝑔 𝐱1 − 𝐱2 (198) 𝑗 𝐱 (𝐱 ) = 𝜆 𝐮𝑊 (𝐱, 𝐮)𝑑𝐮 (206)
( ) ∫
The direction factor 𝑔 𝐱1 − 𝐱2 is actually the spectral CCF of the optical √
field. The intensity of the illumination is perfectly homogeneous. In the 1
𝑗 𝑧 (𝐱 ) = 𝑘2 − 4𝜋 2 |𝐮|2 𝑊 (𝐱, 𝐮)𝑑𝐮 (207)
central differential coordinate system, the CSD becomes 𝑘∫
( ) It can be seen that the transverse flux jx and the longitudinal flux jz cor-
𝐱′ 𝐱′ ( )
𝑊𝜔 𝐱 + , 𝐱 − = 𝑆0 𝑔 𝐱 ′ (199) respond to the integrals of the RHS and LHS of Liouville transport equa-
2 2
tion over the frequency variable 𝐮, respectively. Geometrical vector flux
Its corresponding WDF is is a radiometric quantity, which describes the light energy propagation
𝑊𝜔 (𝐱, 𝐮) = 𝑆0 𝑔̂ (𝐮) (200) in 3D space. In fact, the Liouville transport equation is an expression of
the energy conservation law, which implies the geometrical vector flux
where 𝑔̂ (𝐮) is the Fourier transform of g(x′). It can be seen that the has zero divergence.
WDF of spatially stationary fields is a function only of the spatial fre- √
Under the paraxial approximation 𝑘2 − 4𝜋 2 |𝐮|2 ≈ 𝑘, the Liouville
quency variable 𝐮. It has a form that is similar to the WDF of incoherent
transport equation can be simplified as
light [Eq. (197)] rotated by 90 degrees in phase space. According to the
Fourier transform property of WDF [Eq. (179)], the spatially stationary 𝜕𝑊 (𝐱, 𝐮)
+ 𝜆𝐮∇𝐱 ⋅ 𝑊 (𝐱, 𝐮) = 0 (208)
field and the incoherent field essentially form a Fourier transform pair 𝜕𝑧
(far-field diffraction). The duality between incoherent light and spatially This transport equation can again be solved explicitly, and the solu-
stationary light is, in fact, the Van Cittert-Zernike theorem [Eq. (151)]: tion reads as
the far-field diffraction of an incoherent light source creates a spatially
𝑊𝑧 (𝐱, 𝐮) = 𝑊0 (𝐱 − 𝜆𝑧𝐮, 𝐮) (209)
stationary field.
which is just the Fresnel diffraction property of WDF [Eq. (179)].
3) Quasi-homogeneous field

The CSD of quasi-monochromatic fields satisfying the quasi- 6.2.6. Coherence measurement
homogeneous model can be expressed as [Eq. (162)] For monochromatic coherent fields, the WDF is defined based on the
( ) 2D complex amplitude [Eq. (167)], so the 4D phase-space representation
( ) 𝐱 +𝐱 2 ( )
𝑊𝜔 𝐱1 , 𝐱2 = 𝑆𝜔 1 𝑔 𝐱1 − 𝐱2 (201) is highly redundant. When the 2D intensity and phase distributions are
2
known, the WDF can be directly calculated by Eq. (167). Therefore, for
In the central differential coordinate system, the CSD becomes monochromatic coherent fields, the reconstruction of WDF essentially
( ) boils down to a phase recovery problem, which has been discussed in
𝐱′ 𝐱′ ( )
𝑊𝜔 𝐱 + , 𝐱 − = 𝑆𝜔 (𝐱)𝑔 𝐱′ (202) detail in previous sections.
2 2
It is more complicated when the field is not strictly coherent. Gen-
Such quasi-homogeneous field can be locally considered as spatially erally, the phase space WDF constitutes a rigorous and non-redundant
stationary, having, however, a slowly varying intensity, i.e., S𝜔 is a representation for partially coherent fields. The 2D amplitude and phase
relatively slow-varying signal compared with g. The WDF of quasi- are insufficient to determine a partially coherent field unambiguously.
homogeneous field takes the form The complete characterization of the 4D coherence function (so-called
𝑊𝜔 (𝐱, 𝐮) = 𝑆𝜔 (𝐱)𝑔̂ (𝐮) (203) coherence measurement or coherence retrieval) has always been an ac-
tive research area. To fully characterize 4D partially coherent fields,
Both incoherent fields and spatially stationary fields are special cases of one idea is to measure or reconstruct the two-point correlation func-
quasi-homogeneous fields: for spatial incoherent fields 𝑔̂ (𝐮) = 1 and for tion, such as MCF, CSD, or MI for quasi-monochromatic fields. After the
spatially stationary fields 𝐼 (𝐱) = 𝑆0 . two-point correlation function is obtained, the WDF can be calculated
by definition [Eq. (164)]. For example, MCF can be directly measured
6.2.5. Transport equation of Wigner distribution function through interferometry based on the contrast and displacement (phase)
In Subsections 2.2.1 and Subsection 6.1.3, we learned that the of the fringe pattern in Yang’s double-slit interference experiment, as
propagation of coherent complex fields and MI/CSD both satisfy the illustrated in Fig. 44(a). However, if we want to measure the 4D MCF
Helmholtz equation [Eqs. (2) and (141)]. Under the paraxial approxi- of a 2D optical field, we need to traverse through the entire 4D space
mation, Helmholtz equation can be further reduced to the paraxial wave (2D by 2D), which is extremely time-consuming and inconvenient for
equation [Eq. (4)]. As a new representation of coherent fields and par- practical implementation. Fortunately, in practical measurements, the
tially coherent fields, WDF should also follow certain wave equation principle of shearing interferometry can be used to create two overlap-
describing its propagation characteristics. Starting from the Helmholtz ping copies of the test beam with a transversal, axial, or rotational shear
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 44. Coherent measurements via interferometry. (a) Raster scan measurements based on Yang’s double slit interference experiment; (b) parallel measurements
based on a Sagnic radial shearing interferometer (PBS: Polarizing Beam Splitter).

based on Mach-Zehnder interferometer [324], Sagnic radial shearing in-

terferometer [325,326], or rotational shearing interferometer [327] to
achieve high-efficiency, parallel measurements of a large collection of
point pairs, as illustrated in Fig. 44(b).
Another category of coherence measurement approaches is based on
phase-space measurements, i.e., measuring or recovering the 4D WDF.
The most well-known method is called phase space tomography (PST)
[314,315]. By introducing asymmetric optical elements (e.g., cylindri-
cal lens) into the field to be measured, and then collecting a large num-
ber of intensity distributions of the optical field at various propagation
distances, the rotational projections of WDF at different angles can be
obtained. Finally, the complete 4D WDF can be reconstructed in a way
similar to the conventional CT.
Before introducing PST, let us first review the basic principle of the
conventional CT [328]. For simplicity, we first consider a 2D function
f(x, y), and its projection can be represented as the integration along a
straight line l(s, 𝜃)

𝑅𝑓 (𝑠, 𝜃) = 𝑓 (𝑥, 𝑦)𝛿(𝑥 cos 𝜃 + 𝑦 sin 𝜃 − 𝑠)𝑑 𝑥𝑑 𝑦

∬

= 𝑓 (𝑠 cos 𝜃 + 𝑡 sin 𝜃, 𝑠 sin 𝜃 − 𝑡 cos 𝜃)𝑑𝑡 (210) Fig. 45. Illustration of Radon transform. (a-b) One point in 2D space and the
∫
corresponding Radon transform; (c-d) the projection of a 2D object along a cer-
The line equation l(s, 𝜃) can be expressed as tain direction and the Radon transform of the 2D object.

𝑥 cos 𝜃 + 𝑦 sin 𝜃 = 𝑠 (211)

be written as
or equivalently,
{ 𝜕 𝑅𝑓 (𝑠, 𝜃) 1
𝑥 = 𝑠 cos 𝜃 − 𝑡 sin 𝜃 𝑓 (𝑥, 𝑦) = 𝑑 𝑠𝑑 𝜃 (213)
(212) ∬ 𝜕𝑠 2𝜋 2 (𝑥 cos 𝜃 + 𝑦 sin 𝜃 − 𝑠)
𝑦 = 𝑠 sin 𝜃 + 𝑡 cos 𝜃
Eq. (213) shows that after acquiring the projection Rf (s, 𝜃) at each angle,
where s represents the distance from the origin to the line, and 𝜃 repre- the tomogram of the object can be reconstructed. Such a tomography
sents the positive angle from the line to y-axis (or the positive angle from method is called inverse Radon transform.
the projection plane to the x-axis). When 𝜃 and s are fixed, Eq. (211) rep- The mathematical mechanism of CT can also be understood by rein-
resents a straight line with the angle 𝜃 between y-axis and the distance terpreting the problem in the spatial frequency domain. As the central
s from the origin. The projection of f(x, y) onto the line can be obtained theorem in classical CT, the Fourier slice theorem connects the projection
by Eq. (210). When 𝜃 is fixed and s is a variable, Eq. (211) represents data to the original function in Fourier space, which allows efficient CT
a set of parallel lines with the angle 𝜃. In this case, the 1D projection reconstructions with FFT. Based on the definition of Fourier transform
of the 2D function f(x, y) along the line can be obtained by Eq. (210). and variable substitution, it can be proven that the 1D Fourier trans-
If 𝜃 and s are both variables, the function f(x, y) will be mapped to form 𝑅̂ 𝑓 (𝛾, 𝜃) (𝛾 corresponding to the spatial frequency s coordinate) of
another 2D space (s, 𝜃) to obtain projections in different directions, as the 1D projection at angle 𝜃 is distributed along a straight line at the
illustrated in Fig. 45(c). Note that Rf (s, 𝜃) is not defined in the polar same angle in the Fourier space of the 2D object f(x, y), as illustrated in
coordinate system but on the surface of a semi-cylinder. Spreading the Fig. 46.
semi-cylinder surface onto a plane, we obtain an image as shown in |𝑢=𝛾 cos 𝜃
Fig. 45(d). This image is also called a “Sinogram”, because a point in (x, 𝑅̂ 𝑓 (𝛾, 𝜃) = 𝑓̂(𝑢, 𝑣)| = 𝑓̂(𝛾 cos 𝜃, 𝛾 sin 𝜃) = 𝑓̂(𝛾, 𝜃) (214)
|𝑣=𝛾 sin 𝜃
y) plane will be mapped into a sinusoid, as illustrated in Figs. 45(a) and
where (𝛾, 𝜃) is the polar coordinates in the frequency domain
45(b). The mapping relation expressed in Eq. (210) forms a transform { √
from the Cartesian coordinate (x, y) system into (s, 𝜃) coordinate system, 𝛾 = 𝑢2 + 𝑣2
(215)
which is so-called Radon transform [329]. The inverse of Eq. (210) can 𝜃 = arctan 𝑣𝑢
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 47. Correspondence between Wigner distribution function and Ambiguity

Fig. 46. Radon transform and the spatial and Fourier domain correspondence. function.

This theorem allows to access the whole 2D Fourier space of the ob-
ject by collecting a set projection data Rf (s, 𝜃) at different angles. Once
𝑓̂(𝛾, 𝜃) is completed, the object function can be reconstructed by a simple
2D inverse Fourier transform.
Next, we turn our attention to PST [314,315]. Based on the CT
principle discussed earlier, PST recovers the WDF by successive pro-
jection measurements of the WDF at different rotation angles. Accord-
ing to the properties of WDF introduced in Subsection 6.2.2, for quasi-
monochromatic fields, intensity is just a projection of WDF along x. In
order to rotate the WDF, the simplest way is to take the Fourier trans-
form (by far-field diffraction or introducing a lens) so that the WDF
exchanges the space and spatial frequency variables, thus rotates 𝜋/2
in phase space. The intensity then corresponds to the projection of the
WDF along 𝐮, so the projection at the angle 𝜋/2 is obtained. As repre-
sented by Eq. (150), the CSDs in the focal and back focal planes of a thin
lens form a 4D Fourier transform pair
( ) ( )
𝑊 𝐱1 , 𝐱2 = 𝜆̄21𝑓 2 ∬ 𝑊𝑖𝑛 𝐱′ 1 , 𝐱′ 2
{ [ ( ) ( )]} ′ (216)
2𝜋
̄ 𝐱 2 ⋅ 𝐱2 − 𝐱1 − 𝐱 1 ⋅ 𝐱2 − 𝐱1 𝑑 𝐱 1 𝑑 𝐱′ 2
exp 𝜆𝑓 ′ ′

which can be denoted as Fig. 48. Basic principle of phase space tomography. (a) Vertical projection (0-
( ) ( )| order fractional Fourier transform); (b) quarter rotation projection (1/4-order
𝑊 𝐱1 , 𝐱2 = 𝑊̂ 𝑖𝑛 𝐮1 , 𝐮2 | 𝐱1,2 (217)
|𝐮1,2 =± 𝜆𝑓 fractional Fourier transform); (c) 90-degree rotation projection (1-order frac-
tional Fourier transform); (d) superposition of all projections from different an-
Let 𝑠 = 𝜆𝑓 , according to the Fourier transform property, the WDFs in gles.
the input and output planes have the following relationship
( )
𝑊 (𝐱, 𝐮) = 𝑊𝑖𝑛 −𝑠𝐮, 𝑠−1 𝐱 (218)
by PST is completely consistent with the basic principle of traditional
Therefore, in the normalized coordinates (xn , un ), where 𝐱𝑛 = 𝑠−1∕2 𝐱, CT, as illustrated in Fig. 48.
𝐮𝑛 = 𝑠1∕2 𝐮, the WDF rotates 𝜋/2. The intensities in the front and back In fact, not just limited to the fractional Fourier transform, we can
focal planes of the lens are also use Fresnel diffraction to shear the WDF [332,333]
( )
𝐼 (𝐱 ) = 𝑊 (𝐱, 𝐮)𝑑𝐮 = 𝑊𝑖𝑛 −𝑠𝐮, 𝑠−1 𝐱 𝑑𝐮 (219) 𝑊Δ𝑧 (𝑥, 𝑢) = 𝑊𝑖𝑛 (𝑥 − 𝜆Δ𝑧𝑢, 𝑢) (222)
∫ ∫
Therefore, by introducing a lens, we can obtain the projection of the The corresponding intensity is 𝐼 (𝑥) = ∫ 𝑊𝑖𝑛 (𝑥 − 𝜆Δ𝑧𝑢, 𝑢)𝑑𝑢. Compared
WDF at a different angle. In order to reconstruct the WDF, we need with Eq. (220), it can be found that the projection of WDF at angle 𝜃 is
to collect more projections at various angles. Based on the fractional obtained “equivalently”, and the relation between 𝜃 and the propagation
Fourier transform property of WDF, for 1D optical signals or axisym- distance Δz is
metric optical fields, the corresponding fractional Fourier transform ex- tan 𝜃 = −𝜆Δ𝑧𝑢 (223)
pressed in Eq. (180) directly corresponds to a 𝜃-rotation of the original
WDF Win (x, u) in phase space As shown in Fig. 49, although the transformations of fractional Fourier
transform and Fresnel diffraction in phase space are different, they are
𝑊ℱ𝜃 (𝑥, 𝑢) = 𝑊𝑖𝑛 (𝑥 cos 𝜃 − 𝑢 sin 𝜃, 𝑢 cos 𝜃 + 𝑥 sin 𝜃) (220)
equivalent for obtaining WDF projections. When Eq. (223) is satisfied,
The corresponding intensity signal is just the Radon transform of the the projections obtained by fractional Fourier transform and Fresnel
original WDF WU at angle 𝜃. Since WDF Win (x, u) and AF Ain (u′, x′) diffraction are both along the same line (P1 and P2 in Fig. 49). And the
form a Fourier transform pair [Eq. (169)], the Fourier slice theorem in Fourier transform of the WDF projection obtained by Fresnel diffraction
( )
phase space can be written as [330,331] also corresponds to a 1D slice of the 2D object spectrum 𝐴𝑖𝑛 𝑢′ , −𝜆𝑧𝑢′ .
{ }( )|𝑢′ =𝛾 cos 𝜃 The case we discussed above is limited to 1D signal or 2D axisym-
𝑅̂ 𝑊𝑖𝑛 (𝛾, 𝜃) = ℱ 𝑊𝑖𝑛 (𝑥, 𝑢) 𝑢′ , 𝑥′ | ′ metric fields. However, it should be noted that for a general 2D optical
|𝑥 =𝛾 sin 𝜃
field, its WDF is a 4D function. If we need to reconstruct the 4D WDF
= 𝐴𝑖𝑛 (𝛾 cos 𝜃, 𝛾 sin 𝜃) = 𝐴𝑖𝑛 (𝛾, 𝜃) (221)
based on the CT principle, we need to “rotate” the 4D WDF over two
Fig. 47 illustrates the correspondence between WDF and AF. There- independent phase space planes (𝑥 − 𝑦 and 𝑢𝑥 − 𝑢𝑦 ). However, the frac-
fore, for 1D signals or 2D axisymmetric fields, the reconstruction of WDF tional Fourier transform expressed by Eq. (180) takes the same order on
C. Zuo, J. Li and J. Sun et al.
Fig. 49. Two different transformations of WDF for the implementation of phase
space tomography. (a) WDF of the original signal; (b) WDF after Fresnel diffrac-
tion; (c) WDF after fractional Fourier transform; (d) the correspondence between
Fresnel diffraction and fractional Fourier transform.
Fig. 50. Experimental arrangement of PST for recovering the 4D WDF of the
2D field at the plane of incidence O. A pair of cylindrical lenses oriented per-
pendicularly are used to introduce astigmatism to the measurement.
both axes 𝐱 = (𝑥, 𝑦), i.e., the rotation angles 𝜃 for both axes are identical.
In fact, the fractional Fourier transform represented by the Eq. (180) is
not the general form for 2D functions [310]. The transform kernel of
fractional Fourier transform is a separable function in the Cartesian co-
ordinate system, so its orders (rotation angles) for two axes x, y can as-
sume different values. In optics, to realize asymmetric fractional Fourier
transform for 2D optical fields, astigmatic optics should be introduced
to break this symmetry in x and y dimensions, as illustrated in Fig. 50.
By changing the focal lengths or the distances between the lenses, we
can adjust 𝜃 x , 𝜃 y independently, so as to obtain the 2D rotational projec-
tions of the WDF. The 4D WDF of the input plane can be reconstructed
by inverse Radon transform after the projection data of various angles
are collected.
In addition to these “indirect” measurement methods based on in-
terferometry and PST, we can also employ the property of the WDF as
a “local spectrum (spectrogram)” or generalized radiance to probe the
WDF directly. Two typical approaches are often used.
1) Pinhole scanning: the measured field is spatially localized by an
aperture (usually a pinhole) scanning across x. The corresponding
2D local spectrum can be recorded by far-field diffraction or based
on the Fourier transform property of a thin lens. At a given location
x0 , the captured local spectrum corresponds approximately to a spa-
Optics and Lasers in Engineering 135 (2020) 106187
Fig. 51. Direct phase-space measurement methods. (a) Phase-space measure-
ment based on pinhole scanning; (b) phase-space measurement based on a mi-
crolens array.
tial sampling of the WDF at x0 [334–336]. When the pinhole moves
accross the whole 2D space, the 4D WDF of the input field can be
obtained, as illustrated in Fig. 51(a).
2) Array measurements: the experimental setup is quite similar to the
Shack-Hartmann wavefront sensor [40–42] and light-field camera
[337], as illustrated in Fig. 51(b). This method can be regarded as a
“parallelized” version of the former pinhole scanning method, which
allows single-shot acquisition (the intensity distribution behind each
microlens corresponds to the local spectrum at different positions)
[338,339]. However, the spatial resolution of the measurement is
compromised, which only depends on the pitch of the microlens ar-
ray. The basic principle of these two methods are quite similar: they
both employ the approximate equivalence between WDF and the en-
ergy density function (radiance or light field in computer graphics)
to realize the direct measurement of WDF.
Finally, it should be noted that the equivalence between WDF and
radiance or light field can only be established under certain conditions
(slowly varying coherent field or geometric optics approximation), so
the local spectrum directly measured does not strictly correspond to the
WDF. Strictly speaking, based on the convolution property of WDF, the
spectrogram measured through a pinhole or a lens aperture corresponds
to a smoothed version of the WDF [334,335].
𝑊𝑠 (𝐱, 𝐮) = 𝑊𝑖𝑛 (𝐱, 𝐮) ⊗ 𝑊𝑇 (𝐱, 𝐮) (224)
𝐱,𝐮
where ⊗ represents the 4D convolution over both x and u, Win (x, u) rep-
𝐱,𝐮
resents the WDF of input optical field and WT (x, u) represents the WDF
of aperture function. Although WDF may be negative, the spectrogram
after convolution is always nonnegative and directly measurable. Obvi-
ously, if we want to make the measured spectrogram approach the true
WDF, WT (x, u) should tend to 𝛿(x, u). However, limited by Heisenberg’s
uncertainty principle, it is physically impossible to shrink the support of
WDF in both spatial and spatial-frequency domain. In order to alleviate
this problem, a compromise between the blurring effects in the spatial
and the spatial frequency domains should be made by appropriately de-
signing the aperture function [335,336,340]. An alternative approach
is to compensate for the aperture-induced blurring effect based on the
WDF deconvolution method [341].
6.3. Image formation model under partially coherent illumination
In Section 4, we discussed the imaging formation for coherent imag-
ing systems, where the source at the condenser aperture plane is as-
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 52. Schematic diagram of a microscopic imaging system with partially incoherent illuminations. The condenser diaphragm can be regarded as an incoherent
extended source with a finite size.

sumed to be an ideal on-axis point radiator with strict monochromatic- the objective pupil is multiplied twice by the coherent transfer function
ity. For a practical optical microscopic imaging system, this condition is H(u) [Eq. (63)], and then followed by an inverse transform to finally
difficult to meet because the light source always has a certain band- obtain the mutual intensity in the image plane
width, and the condenser aperture always has a finite size. In other
( ) ( ) ( ) ( )
words, the illumination produced by the light source is neither perfect 𝑊𝐼 𝐱1 , 𝐱2 = 𝑊̂ 𝑂 𝐮1 , 𝐮2 𝐻 𝐮1 𝐻 * 𝐮2 𝑒𝑗2𝜋 (𝐮1 𝐱1 +𝐮2 𝐱2 ) 𝑑 𝐮1 𝑑 𝐮2 (227)
∬
in temporal coherence nor spatial coherence. So in this section, we ex-
amine the effects of partial coherence on microscopic imaging. First, we where H(u1 )H∗ (u2 ) is the mutual coherent transfer function (discussed
ignore the limited aperture effect and discuss the influences of temporal in Subsection 6.1.3). Eq. (227) suggests that for partially coherent imag-
coherence and spatial coherence on image formation. Then we take into ing, the CSD is transferred linearly with the mutual coherent trans-
account the aperture effect and derive the image formation model for fer function. It can also be written as a two-fold convolution of the
a practical microscopic imaging system under partially coherent illumi- sample CSD with the mutual PSF, defined as h(x1 )h∗ (x2 ) (discussed in
nations. Subsection 6.1.3)
( ) ( ) ( ) ( )
6.3.1. Ideal imaging model under partially incoherent illumination 𝑊𝐼 𝐱1 , 𝐱2 = 𝑊𝑂 𝐱 ′ 1 , 𝐱 ′ 2 ℎ 𝐱1 − 𝐱 ′ 1 ℎ∗ 𝐱2 − 𝐱 ′ 2 𝑑 𝐱 ′ 1 𝑑 𝐱 ′ 2 (228)
∬
In this section, we extend the coherent imaging condition discussed
in Subsection 4.2, considering that the condenser aperture becomes in- This suggests that the CSD in the object space is blurred by the mutual
coherent extended source with a finite size [Eq. (145)]. Fig. 52 illus- PSF in both coordinates (𝐱𝟏 and 𝐱𝟐 ). The image intensity I(x) is given by
trates the corresponding optical configuration, where the light source the values “on the diagonal” of the CSD
(condenser aperture diaphragm) is placed at the front focal plane of the ( ) ( ) ( ) ( ) ( )
condenser lens, producing a highly homogeneous illumination field in 𝐼 (𝐱) = 𝑊𝐼 (𝐱, 𝐱) = 𝑊𝑆 𝐱1 , 𝐱2 𝑇 𝐱1 𝑇 * 𝐱2 ℎ 𝐱 − 𝐱1 ℎ∗ 𝐱 − 𝐱2 𝑑 𝐱1 𝑑 𝐱2 (229)
∬
the sample plane. Based on the Van Cittert-Zernike theorem, the CSD of
the illumination just before the sample is given by the Fourier transform Substituting Eq. (225) into Eq. (229), the expression for the intensity
of the source intensity distribution S(u) [Eq. (151)] can be simplified consequently

( ) ( ) | ( ) ( ) |2
𝑆 (𝐮)|| 𝑇 𝐱′ ℎ 𝐱 − 𝐱′ 𝑒𝑗2𝜋𝐮𝐱 𝑑 𝐱′ || 𝑑𝐮 ≡
′
𝑊𝑆 𝐱1 , 𝐱2 = 𝑊𝑆 𝐱1 − 𝐱2 = 𝑆 (𝐮)𝑒𝑗2𝜋𝐮⋅(𝐱1 −𝐱2 ) 𝑑𝐮 (225) 𝐼 (𝐱 ) = 𝑆 (𝐮)𝐼𝐮 (𝐱)𝑑𝐮
∬ ∫ |∫ | ∫
(230)
It only depends on the coordinate difference between the two points in
the plane. In the following of this subsection, the formal description of Eq. (230) suggests that the intensity captured at the image plane can
the imaging process is limited to quasi-monochromatic light and unit be interpreted as an incoherent superposition of the coherent partial
(1.0 × ) magnification. An extension to polychromatic illuminations and images Iu (x) arising from all points of the incoherent source.
arbitrary magnification is straightforward.
When the illumination light transmits a thin object of complex am- 6.3.2. Effect of spatial coherence on image formation
plitude 𝑇 (𝐱) = 𝑎(𝐱) exp [𝑗𝜙(𝐱)], the CSD just behind the object is given by In Subsection 4.2, we considered the case of plane wave coherent
illumination, i.e., the light source at the condenser aperture plane is an
( ) ( ) ( ) ( )
𝑊𝑂 𝐱1 , 𝐱2 = 𝑊𝑆 𝐱1 , 𝐱2 𝑇 𝐱1 𝑇 * 𝐱2 (226) on-axis point radiator 𝑆 (𝐮) = 𝛿(𝐮). For such a coherent imaging case, the
intensity in the image plane can be represented as [Eq. (60)]
The CSD now depends on the complex amplitudes of the sample and the
| ( ) ( ) |2
CSD of the illumination. The modulation of the optical field by the imag- 𝐼𝟎 (𝐱) = || 𝑇 𝐱′ ℎ 𝐱 − 𝐱′ 𝑑 𝐱′ || =|𝑇 (𝐱) ⊗ ℎ(𝐱)|2 (231)
ing system can be described in the spatial frequency domain based on |∫ |
the transfer function theory, as discussed in Subsection 5.1.3. After the Let us consider another extreme case: the source is infinitely ex-
transmitted optical field passing through the objective lens, the CSD in tended such that the CSD of the illumination field degenerates to a 𝛿
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

function.
( ) ( )
𝑊𝑆 𝐱1 , 𝐱2 = 𝐶 𝑒𝑗2𝜋𝐮⋅(𝐱1 −𝐱2 ) 𝑑𝐮 = 𝐶𝛿 𝐱1 − 𝐱2 (232)
∬
Substituting Eq. (232) into Eq. (229) and neglecting the unimportant
constant factor, we obtain
( ) ( ) ( ) ( ) ( )
𝐼 (𝐱 ) = 𝛿 𝐱1 − 𝐱2 𝑇 𝐱1 𝑇 * 𝐱2 ℎ 𝐱 − 𝐱1 ℎ∗ 𝐱 − 𝐱2 𝑑 𝐱1 𝑑 𝐱2
∬
=|𝑇 (𝐱)|2 ⊗ |ℎ(𝐱)|2 (233)
Eqs. (232) and (233) represent two extreme cases for partially coher-
ent imaging: completely coherent and incoherent. For partially coherent
imaging, it should be somewhere in between: the extended source has
a finite size. In order to reveal the influence of spatial coherence on the
formed intensity image more clearly, the limited-aperture effect of the
objective lens is neglected [the NA of the objective lens is infinite, i.e.,
P(u) ≡ 1 in Eq. (63)]. When the object is imaged at a defocus distance
Δz, the coherent transfer function degenerates into the angular spec-
trum transfer function for free-space propagation [Eq. (30)]. Under the Fig. 53. Blurring effect due to spatial coherence of the light source [342]. (a)
paraxial approximation, the impulse response function can be further An off-axis point source produces tilted plane wave that travels with transverse
simplified as [Eq. (17)] vector direction −𝐬∕𝑓 ; (b) intensity blurring resulting from the finite source size
{ } under defocus. Rays passing through a single point on the sample will produce
1 𝑗𝜋
ℎΔ𝑧 (𝐱) = exp(𝑗 𝑘Δ𝑧) exp |𝐱|2 (234) a circular spot in the defocused planes located at ± Δz of diameter 2Δzs/f.
𝑗𝜆Δ𝑧 𝜆Δ𝑧
Substituting Eq. (234) into Eqs. (230) and (231), and ignoring the
constant and phase factors that are independent of the transverse inten- 6.3.3. Effect of temporal coherence on image formation
sity The temporal coherence of the light source is another important
{ ] } |2
𝑗𝜋 [ 2 | ′ |2
| ( ) factor affecting the image formation. In current microscopes, the light
| |
𝐼 (𝐱) = 𝑆 (𝐮)| 𝑇 𝐱′ exp |𝐱| + |𝐱 | − 2(𝐱 − 𝜆Δ𝑧𝐮) ⋅ 𝐱′ 𝑑 𝐱′ | 𝑑𝐮 (235) source is usually a halogen lamp or a continuous spectrum LED. Neither
∫ |∫ 𝜆Δ𝑧 |
| |
of them can provide monochromatic illuminations like a laser so that
{ ]} |2
|
| ( ) 𝑗𝜋 [ 2 | ′ |2 | the monochromatic assumption is no longer valid. As we know from
𝐼0 (𝐱) = | 𝑇 𝐱′ exp |𝐱| +|𝐱 | − 2𝐱 ⋅ 𝐱′ 𝑑 𝐱′ | (236) Subsection 6.1, the temporal coherence of a light source depends on its
|∫ 𝜆Δ𝑧 |
| | PSD (spectral distribution). Assuming that the PSD of the light source is
Adding an insignificant constant factor to the intensity in S𝜔 (u), since different wavelengths are uncorrelated, the total intensity
Eq. (235) and comparing it with Eq. (236) gives: distribution of the source is the integral of the PSD over all wavelengths
{ ]} |2
|
| ( ) 𝑗𝜋 [ |
|𝐱 − 𝜆Δ𝑧𝐮|2 +||𝐱′ || − 2(𝐱 − 𝜆Δ𝑧𝐮) ⋅ 𝐱′ 𝑑 𝐱′ | 𝑑𝐮
2
𝐼 (𝐱 ) = 𝑆 (𝐮)| 𝑇 𝐱′ exp
∫ |∫
| 𝜆Δ𝑧 |
| 𝐼 𝑆 (𝐮 ) = 𝑆𝜔 (𝐮)𝑑𝜔 (238)
∫
| { [ ] } |2
| ( ′) 𝑗𝜋 | |
𝐱 − 𝜆Δ𝑧𝐮 − 𝐱′ ||
2
= 𝑆 (𝐮)| 𝑇 𝐱 exp 𝑑 𝐱′ | 𝑑𝐮 In the previous subsection, the light source is assumed to be a quasi-
∫ |∫
| 𝜆Δ𝑧 | |
|
𝐱 ( ) monochromatic extended source with an intensity distribution of S(u).
𝐮= 𝜆𝑓
Δ𝑧𝐱 In this section, we assume that the light source has a certain spectral
= 𝑆 (𝐮)𝐼𝟎 (𝐱 − 𝜆Δ𝑧𝐮)𝑑𝐮 = 𝑆 ⊗ 𝐼 𝟎 (𝐱 ) (237)
∫ 𝑓
bandwidth with a PSD of S𝜔 (u). In this case, Eqs. (229) and (230) can
where S(Δzx/f) is the effective intensity distribution of the light source, be rewritten as
and the size is proportional to the defocus distance Δz. The physical
meaning of Eq. (237) can be explained by examining Fig. 53: when the 𝐼 (𝐱 ) = 𝑊𝜔 (𝐱, 𝐱)𝑑𝜔
∫
object is illuminated by an oblique plane wave emitted from an off-
( ) ( ) ( ) ( ) ( )
axis point source, the image will be laterally shifted with a distance = 𝑊𝑆𝜔 𝐱1 , 𝐱2 𝑇𝜔 𝐱1 𝑇𝜔∗ 𝐱2 ℎ𝜔 𝐱 − 𝐱1 ℎ∗𝜔 𝐱 − 𝐱2 𝑑 𝐱1 𝑑 𝐱2 𝑑𝜔
Δzs/f that is proportional to the defocus distance Δz. Therefore, for an ∫∬
extended incoherent source (e.g., a homogeneous disk with a radius s, (239)
with the increase in Δz, the diffuse spot (defocused PSF) expands pro-
and
portionally, creating a homogeneous circular disc with a radius of Δzs/f.
The diffuse spots produced by different object points are superimposed, | ( ) ( ) |2
𝑆𝜔 (𝐮)|| 𝑇𝜔 𝐱′ ℎ𝜔 𝐱 − 𝐱′ 𝑒𝑗2𝜋𝐮𝐱 𝑑 𝐱′ || 𝑑𝐮𝑑𝜔
′
resulting in a blurred diffraction pattern (defocused intensity image). 𝐼 (𝐱 ) =
∬ | ∫ |
In order to alleviate the blurring effect induced by the ex-
tended source, we can reduce the defocus distance Δz. When Δz → 0, ≡ 𝑆𝜔 (𝐮)𝐼𝐮,𝜔 (𝐱)𝑑 𝐮𝑑 𝜔 (240)
∬
S(Δzx/f) → 𝛿(u), and there will be no source blur. We can obtain an
intensity distribution I0 (x) that is completely consistent with the co- Eqs. (239) and (240) indicate that the intensity distribution of the
herent case. However, as discussed in Subsection 4.2, the premise of image is the superposition of the power density produced by different
TIE phase retrieval is to create phase contrast by defocusing, thus con- wavelength components. This conclusion seems straightforward because
vert phase information into intensity information. The phase compo- it only needs to integrate over all different wavelengths. However, it
nent of the object cannot be visualized without defocusing. Therefore, should be stressed that both the complex transmittance of the object and
it is not difficult to understand that using extended sources (open up the PSF of the imaging system in Eqs. (239) and (240) are wavelength-
the condenser diaphragm) will attenuate the phase contrast in a micro- dependent functions in a strict sense. More specifically, the wavelength-
scopic image system. Note that the precondition for the establishment of dependent RI induced sample dispersion and chromatic aberration of-
Eq. (237) is that the aperture of the imaging system is infinite [𝑃 (𝐮) ≡ 1]. ten complicate accurate phase determination. Therefore, in practical
Eq. (237) may break down when considering the finite NA of the objec- phase imaging systems, the sample is often illuminated with quasi-
tive (see Subsection 6.3.4 for detailed discussions). monochromatic illumination by introducing a narrow bandpass filter
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

into the illumination path or using monochromatic LEDs as illumina-

tion sources. When it is necessary to use polychromatic illuminations,
we have to impose certain assumptions on the sample: e.g., it should
be non-dispersive and the absorption is wavelength-independent within
the spectral range of the illumination (which are generally valid for un-
stained cells).
To better explain the influence of temporal coherence on image for-
mation, let us turn our attention back to the complex impulse response
function under Fresnel approximation
{ }
1 𝑗𝜋
ℎΔ𝑧 (𝐱) = exp(𝑗 𝑘Δ𝑧) exp |𝐱|2 (241)
𝑗𝜆Δ𝑧 𝜆Δ𝑧
It is obvious that the impulse response function is related to the illu-
mination wavelength. The first two terms on the RHS of Eq. (241) are
constant and phase factors{ that are} independent of the lateral coordi-
𝑗𝜋
nates. The third term exp 𝜆Δ𝑧
|𝐱|2 is a quadratic phase factor which
determines the intensity distribution. Close inspections reveal that the
wavelength 𝜆 and Δz appear in pair, which suggests that a change in Fig. 54. Geometric interpretation of the transmission cross coefficient (TCC).
wavelength (𝜆 → 𝜆 + 𝛿𝜆) is equivalent to a change in propagation dis-
tance (Δ𝑧 → Δ𝑧 + 𝛿𝑧), where
𝛿𝜆Δ𝑧 The image intensity is therefore computed by the linear superposition
𝛿𝑧 = − (242) of the interference patterns of plane waves with spatial frequencies u1
𝜆 + 𝛿𝜆
Therefore, if polychromatic illuminations are used, the intensity fi- and 𝐮2 [𝑒𝑗2𝜋𝐱(𝐮1 −𝐮2 ) ], and amplitude given by the TCC and the object
nally captured at the defocused plane can be regarded as the superpo- spectrum at u1 and u2 [T(u1 )T∗ (u2 )]. In essence, TCC is a 2D convolution
sition of many diffraction patterns of different wavelengths (or equiv- integral, which can be calculated by integrating the three functions over
alently, at different defocus distances). Obviously, the superposition of the overlapping area of the light source and the two shifted complex
mutually uncorrelated diffuse spots (defocused PSFs) at different propa- pupils, as illustrated in Fig. 54. The result of the integral at a given
gation distances results in a blurred diffraction pattern. Therefore, sim- position (ux1 , uy1 ) and (ux2 , uy2 ) corresponds to one element in the 4D
ilar to the lateral blurring effect due to the spatial coherence, the tem- TCC matrix. When the two pupil functions shift across the (ux , uy ) plane,
poral coherence induced blurring results from the superposition of dif- the 4D TCC matrix can be constructed.
ferent diffuse spots at different axial propagation distances. Finally, it The image formation in partially coherent systems is not linear in
is worth noting that although the use of polychromatic illuminations either amplitude or intensity but is bilinear, which makes image restora-
attenuates the phase contrast in the defocused plane, the wavelength- tion and phase recovery more complicated. To simplify the mathemat-
dependent nature of the Fresnel diffraction can be wisely utilized to get ical formulation, the weak object approximation is often assumed to
rid of mechanical defocusing in TIE phase retrieval [191,212]. linearize the phase retrieval problem, which was discussed under the
condition of coherent imaging (see Subsection 4.3.1). In the following,
6.3.4. Transmission cross-coefficient (TCC) and weak object transfer we discuss the weak object approximation for partially coherent imag-
function (WOTF) ing. The complex transmittance of a weak object can be represented as
In general, the image formation in microscopic imaging systems can [Eq. (77)]
be described by Fourier transform and linear filtering: a coherent imag- 𝜙(𝐱)≪1
ing system is linear in complex amplitude [Eq. (231)], while an inco- 𝑇 (𝐱) = 𝑎(𝐱) exp [𝑗𝜙(𝐱)] ≈ [𝑎0 + Δ𝑎(𝐱)][1 + 𝑗𝜙(𝐱)]
herent imaging system is linear in intensity [Eq. (232)]. However, for Δ𝑎(𝐱)≪𝑎0
≈ 𝑎0 + Δ𝑎(𝐱) + 𝑗 𝑎0 𝜙(𝐱) (246)
partially coherent imaging, the image formation is more complicated
due to the nonlinear dependency of the image intensity on the object The corresponding Fourier transform can be expressed as
transmittance, light source, and imaging system. In Subsection 6.3.1, we [ ]
𝑇̂ (𝐮) = 𝑎0 𝛿(𝐮)+𝜂( ̂ 𝐮)
̂ 𝐮) + 𝑗 𝜙( (247)
gave the representations of the intensity distribution under partially co-
herent illuminations. Eq. (230) has a clear physical meaning: a partially Then the mutual spectrum of the object can be written as
coherent image can be regarded as the superposition of sub-images gen- ( ) ( ) ( )[ ( ) ( )] ( )[ ( ) ( )]
𝑇̂ (𝐮1 )𝑇̂ ∗ (𝐮2 ) = 𝑎20 𝛿 𝐮1 𝛿 𝐮2 + 𝑎0 𝛿 𝐮2 𝜂̂ 𝐮1 + 𝑗 𝜙̂ 𝐮1 + 𝑎0 𝛿 𝐮1 𝜂̂ 𝐮2 + 𝑗 𝜙̂ 𝐮2
erated by different point sources. The image formation model described
by Eq. (230) is the so-called Abbe’s superposition method, which has a (248)
clear physical meaning: a partially coherent image can be regarded as Noted that the high order interference terms are neglected because the
the superposition of sub-images generated by different source points in scattered light is weak, which just corresponds to the first order Born
the condenser aperture plane. It should be noted that the partially co- approximation commonly used in diffraction tomography (discussed in
herent imaging theory of a microscope was first established by Hopkins Section 8). Substituting Eq. (248) into Eq. (245) yields the intensity of
[247] from a different perspective in 1950s. Starting from Eq. (229) and the partially coherent image
combining Eq. (225), we can derive the following Fourier integral { }
[ ]
𝐼 (𝐱) = 𝑎20 𝑇 𝐶 𝐶 (𝟎, 𝟎)+2𝑎0 Re 𝑇 𝐶 𝐶 (𝐮, 𝟎) 𝜂( ̂ 𝐮) 𝑒𝑗2𝜋𝐱𝐮 𝑑𝐮
̂ 𝐮)+𝑗 𝜙(
𝐼(𝐱) = 𝑆(𝐮)𝑇̂ (𝐮1 )𝑇̂ ∗ (𝐮2 )𝐻(𝐮 + 𝐮1 )𝐻 ∗ (𝐮 + 𝐮2 )𝑒𝑗2𝜋𝐱(𝐮1 −𝐮2 ) 𝑑 𝐮1 𝑑 𝐮2 𝑑𝐮 (243) ∫
∭
(249)
Separating the contributions of the specimen and imaging system leads where the simple relation 𝑇 𝐶 𝐶 ∗ (𝟎, 𝐮)=𝑇 𝐶 𝐶 (𝐮, 𝟎)
is used (TCC is Hermi-
to the notion of the transmission cross-coefficient (TCC) tian symmetric). It now becomes obvious that the image contrast due to
( ) the absorption and phase are decoupled and linearized. In the follow-
𝑇 𝐶 𝐶 𝐮1 , 𝐮2 = 𝑆(𝐮)𝐻(𝐮 + 𝐮1 )𝐻 ∗ (𝐮 + 𝐮2 )𝑑𝐮 (244)
∬ ing, we denote the TCC(u, 0), the linear part of the TCC, as weak object
With the TCC, we can simplify Eq. (243) as follows transfer function (WOTF)
( ) ( ) ( ) ( )
𝐼(𝐱) = 𝑇̂ (𝐮1 )𝑇̂ ∗ (𝐮2 )𝑇 𝐶 𝐶 𝐮1 , 𝐮2 𝑒𝑗2𝜋𝐱(𝐮1 −𝐮2 ) 𝑑 𝐮1 𝑑 𝐮2 (245) 𝑊 𝑂𝑇 𝐹 (𝐮) ≡ 𝑇 𝐶 𝐶 (𝐮, 0) = 𝑆 𝐮′ 𝐻 𝐮′ +𝐮 𝐻 * 𝐮′ 𝑑 𝐮′ (250)
∬ ∬
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

illumination or detection schemes are widely adopted in phase-contrast

imaging, wavefront sensing, and QPI, e.g., differential phase-contrast
(DPC) microscopy [343–345], pyramid wavefront sensor [44,346], and
partitioned/programmable aperture microscopy (PAM) [347,348].
The situation discussed above is limited to an aberration-free optical
system. In fact, another approach to break the realness of WOTF is to
introduce aberration. Optical aberration produces an additional imagi-
nary component in the WOTF, resulting in phase contrast. Defocusing is
the most convenient way to introduce aberration into the optical system.
When the imaging system is defocused with a distance Δz, the coherent
transfer function becomes the product of the objective pupil function
P(u) and the angular spectrum transfer function [Eq. (30)]
√
2
𝐻 (𝐮) = 𝑃 (𝐮)𝐻Δ𝑧 (𝐮) = 𝑃 (𝐮)𝑒𝑗𝑘Δ𝑧 1−𝜆 |𝐮|
2
(255)
Substituting the complex pupil function into Eq. (250) results in a com-
plex and even WOTF

𝑊 𝑂𝑇 𝐹 (𝐮) ≡ 𝑇 𝐶 𝐶 (𝐮, 0)
( √ √ )
Fig. 55. Geometric interpretation of WOTF for three different illuminations. (a) ( ) ( ) ( ) 𝑗𝑘Δ𝑧 − 1−𝜆2 |𝐮′ |2 + 1−𝜆2 |𝐮+𝐮′ |2
= 𝑆 𝐮′ 𝑃 𝐮′ 𝑃 𝐮′ + 𝐮 𝑒 𝑑 𝐮′
Incoherent imaging (light source larger than or equal to the objective pupil); (b) ∬
for partially coherent imaging (light source smaller than the objective pupil); (c) (256)
partially coherent oblique illumination.
Similarly, according to the symmetries of Fourier transform, the transfer
functions for the absorption and phase components of a weak object are
Since the absorption and phase of the object are both real functions, then given by the real and imaginary components of the WOTF
according to the properties of Fourier transform, the real part of 𝜂(
̂ 𝐮) is
an even function, and the imaginary part of 𝜂(
̂ 𝐮) is an odd function. Due 𝐻𝐴 (𝐮)= Re [𝑊 𝑂𝑇 𝐹 (𝐮)] (257)
to the imaginary unit before the phase spectrum, the real part of 𝑗 𝜙( ̂ 𝐮)
̂ 𝐮) is an even function.
is an odd function, and the imaginary part of 𝑗 𝜙( 𝐻𝑃 (𝐮) = − Im [𝑊 𝑂𝑇 𝐹 (𝐮)] (258)
For an aberration-free microscopic system with axisymmetric illumi-
nation and objective pupil, the coherent transfer function is the objec- 6.3.5. Phase transfer functions for coherent, incoherent, and partially
tive pupil function 𝐻 (𝐮) = 𝑃 (𝐮). For bright-field imaging, the circular coherent imaging
condenser aperture diaphragm is enclosed by the objective pupil In Section 4, we know that the spatial coherence of the illumination
( ) ( ) in a microscope can be quantified by coherence parameter. The coher-
𝑊 𝑂𝑇 𝐹 (𝐮) ≡ 𝑇 𝐶 𝐶 (𝐮, 0) = 𝑃 𝐮 ′ 𝑃 𝐮 ′ +𝐮 𝑑 𝐮 ′ (251)
∬ ence parameter is defined as the ratio of the illumination NA to the
objective NA [177,178,246,249], which is equivalent to the radius ratio
Now the WOTF(u) is always real and even, as illustrated in Fig. 55. When
̂ 𝐮), the real part of the product of the condenser aperture to the objective aperture
the real WOTF(u) is multiplied with 𝑗 𝜙(
becomes an odd function, and the imaginary part of the product becomes 𝑁 𝐴𝑖𝑙𝑙
𝑠= (259)
an even function. According to the symmetries of Fourier transform, the 𝑁 𝐴𝑜𝑏𝑗
Fourier transform of the odd, real function is an odd, imaginary func- Under coherent illuminations, the source degenerates into the ideal on-
tion, and the Fourier transform of an even, imaginary function is still an axis point radiator 𝑆 (𝐮)=𝛿(𝐮), i.e., s → 0. Then, the partially coherent
even, imaginary function. Therefore, the phase component is a purely WOTF degenerates into the complex pupil function
imaginary function after Fourier transform, so it cannot be reflected in (√ )
the intensity image. Therefore, if the microscope is in-focus, the phase 𝑗𝑘Δ𝑧 1−𝜆2 |𝐮|2 −1
does not manifest itself in the intensity image. In the same way, it can 𝑊 𝑂𝑇 𝐹𝑐𝑜ℎ (𝐮) = 𝑃 (𝐮)𝑒 (260)
be found that the absorption component is a purely real function after and the real and the imaginary parts of the WOTF correspond to the ATF
Fourier transform, so only the absorption component can be observed in and PTF, respectively
the in-focus image. For a general partially coherent imaging system, the [ (√ )]
odd-symmetric and even-symmetric components of the WOTF(u) corre- 𝐻𝐴 (𝐮)=𝑃 (𝐮) cos 𝑘Δ𝑧 1 − 𝜆2 |𝐮|2 − 1 (261)
spond to the contributions of amplitude and phase components in the
intensity signal, respectively. We define them as the ATF and PTF for [ (√ )]
partially coherent imaging 𝐻𝑃 (𝐮) = −𝑃 (𝐮) sin 𝑘Δ𝑧 1 − 𝜆2 |𝐮|2 − 1 (262)
{ [ ]} (√ )
̂ 𝐮) + 𝐻𝑃 (𝐮)𝜙(
𝐼 (𝐱) = 𝑎20 𝑇 𝐶 𝐶 (𝟎, 𝟎) + 2𝑎0 𝑅𝑒 ℱ −1 𝐻𝐴 (𝐮)𝜂( ̂ 𝐮) (252)
Under the paraxial approximation, 𝑗𝑘Δ𝑧 1 − 𝜆2 |𝐮|2 − 1 ≈
where −𝑗𝜋𝜆Δ𝑧|𝐮|2 , and the transfer functions can be simplified as
* ( )
𝐻𝐴 (𝐮) = 𝑊 𝑂𝑇 𝐹 (𝐮) + 𝑊 𝑂𝑇 𝐹 (−𝐮) (253) 𝐻𝐴 (𝐮) = 𝑃 (𝐮) cos 𝜋𝜆Δ𝑧|𝐮|2 (263)

( )
𝐻𝑃 (𝐮) = 𝑊 𝑂𝑇 𝐹 * (𝐮) − 𝑊 𝑂𝑇 𝐹 (−𝐮) (254) 𝐻𝑃 (𝐮) = 𝑃 (𝐮) sin 𝜋𝜆Δ𝑧|𝐮|2 (264)

corresponding to the partially coherent ATF and partially coherent PTF.
Therefore, in order to generate phase contrast signal, the symmetry of
either source or pupil function has to be broken. If the illumination or
the pupil function is asymmetric, the PTF HP (u) will be non-zero so
that the phase component can manifest in the image. Such asymmetric
These expressions are identical to Eqs. (92) and (93) discussed in
Section 4. For weak defocusing Δz → 0, the sine term can be further
approximated by a parabolic function sin (𝜋𝜆Δz|u|2 ) ≈ 𝜋𝜆Δz|u|2 , which
is a Laplacian in the Fourier space, corresponding to the PTF implied in
TIE.
C. Zuo, J. Li and J. Sun et al.
Fig. 56. Transfer functions of a rotation-symmetric system. (a) CTF in the co-
herent case; (b) OTF in the incoherent case.
Next we discuss the other extreme case, incoherent illumination.
Strictly speaking, spatially incoherent illumination requires an infinitely
extended source, i.e., the coherence parameter s → ∞ and the CSD re-
duces to a 𝛿 function [Eq. (232)]. However, when considering a weak
object, as the illumination NA increases to the objective NA or even
larger, 𝑠 = 𝑁 𝐴𝑖𝑙𝑙 ∕𝑁 𝐴𝑜𝑏𝑗 ≥ 1, the circular condenser aperture is limited
by the objective pupil. The form of WOTF will no longer change and
become the autocorrelation of the coherent transfer function
( ) ( )
𝑊 𝑂𝑇 𝐹𝑖𝑛𝑐𝑜ℎ (𝐮) = 𝐻 𝐮′ +𝐮 𝐻 * 𝐮′ 𝑑 𝐮′ (265)
∬
At this time the system behaves the same as an incoherent microscope
[142,178,249] [but note that it is not really incoherent (which requires
s → ∞) because a weak object is assumed in this case]. For an axisym-
metric optical imaging system, the coherent transfer function H(u) is an
even function of u, then WOTFincoh (u) is always real (even under de-
focusing), and its PTF is always 0. Hence the phase component of the
specimen cannot be imaged. Consequently, it is impossible to perform
phase retrieval under incoherent illumination.
For an ideal in-focus axisymmetric system with circular apertures
(no pupil aberrations), the WOTF becomes a convolution of two circles
of equal size (as shown in Fig. 56), which can be represented as
√
⎡ ( ) ( )2 ⎤
2⎢ 𝜌̄ 𝜌̄ 𝜌̄ ⎥
𝑊 𝑂𝑇 𝐹𝑖𝑛𝑐𝑜ℎ (𝜌)
̄ = arccos − 1− 𝜌̄ < 2 (266)
𝜋⎢ 2 2 2 ⎥
⎣ ⎦
where 𝜌̄ is the normalized radical spatial coordinate with respect to the
coherent resolution limit NAobj /𝜆
𝜆 𝐻 (𝜌)
𝜌̄ = 𝜌 (267)
𝑁 𝐴𝑜𝑏𝑗
After normalization, the coherent diffraction limit NAobj /𝜆 corresponds
to 𝜌̄ = 1. Fig. 57(a) shows the in-focus 𝑊 𝑂𝑇 𝐹 𝑖𝑛𝑐𝑜ℎ (𝜌)
̄ , which is purely
real-valued and can only image the absorption part. It can be seen that
the absorption contrast gradually reduces as the increase in spatial fre-
quency, with a 61% relative contrast loss at the coherent resolution limit
(NAobj /𝜆), and finally diminishes to zero at the incoherent diffraction
limit (2NAobj /𝜆). The contrast loss is even more severe when the system
is defocused, as demonstrated by the 𝑊 𝑂𝑇 𝐹 𝑖𝑛𝑐𝑜ℎ shown in Fig. 57(b).
From the above analysis, we know that under coherent illumination,
phase contrast can be produced by defocusing, but imaging resolution is
restricted to the coherent diffraction limit. Incoherent illumination pro-
vides higher imaging resolution (twice the coherent diffraction limit),
but the phase component of the sample cannot be imaged. A compromise
between the phase contrast and imaging resolution can be achieved by
using partially incoherent illuminations with a reduced source dimen-
sion (0 < s < 1). Figs. 58 and 59 show the partially coherent ATF and PTF
Optics and Lasers in Engineering 135 (2020) 106187
for various coherent parameters and defocus distances. These results are
numerically calculated from Eq. (250) directly, as a weighted area of
( )
overlap over H∗ (u′) and 𝐻 𝐮′ +𝐮 . For an ideal in-focus axisymmetric
system with circular apertures, the WOTF can be represented analyti-
cally, which is just the overlapping area of two circles with different
radii
⎧𝜋𝑠2 0 ≤ 𝜌̄ ≤ 1 − 𝑠
⎪ ( ) ( )√ ( 2 2 )2
⎪𝑠2 arccos 𝜌̄2 +𝑠2 −1 − 𝜌̄2 +𝑠2 −1 𝑠2 − 𝜌̄ +2𝑠𝜌̄ −1
𝑊 𝑂𝑇 𝐹𝑠 (𝜌)
̄ =⎨ 2𝜌𝑠
̄ 2𝜌̄
( 2 2 ) ( 2 2 ) √ ( 2 2 )2
⎪
⎪+𝑠2 arccos 𝜌̄ −𝑠 +1 − 𝜌̄ −𝑠 +1 1 − 𝜌̄ −2𝑠𝜌̄ +1 1 − 𝑠 ≤ 𝜌̄ ≤ 1 + 𝑠
⎩ 2𝜌𝑠
̄ 2𝜌̄
(268)
Fig. 57 gives the corresponding in-focus WOTFs for different co-
herent parameters. As can be seen, the cut-off frequency is extended
to 1 + 𝑠, which is resolution limit for partially coherent imaging. Al-
though incoherent illumination has a higher resolution limit, details at
a higher frequency are imaged with low contrast. On the other hand, for
partially coherent imaging, a higher image contrast is achieved due to
the better transfer at the low frequencies. For normal bright-field imag-
ing (for absorptive sample), it has transpired to be a good compromise
when the light source image fills approximately half area of the pupil
(𝑠 = 0.7 − 0.8) for achieving both high imaging resolution and high im-
age contrast (especially for low spatial frequencies).
Next, we focus on the partially coherent PTFs under defocusing. It
can be seen from Fig. 58 that increasing coherence parameter dimin-
ishes the phase contrast considerably. This is in consistent with the well-
known phenomenon that the condenser of a microscope must be stopped
down to produce an appreciable contrast for phase information. But in
this case, the associated reduction in resolution limit becomes a major
issue, leading to a compromised imaging resolution. This suggests that
the coherence parameter plays a very important role in phase imaging.
Closing down the condenser diaphragm can improve the phase contrast
at the expense of a reduced resolution limit. While opening up the con-
denser diaphragm improves the resolution limit, however, the response
of the PTF is significantly attenuated. Therefore, in order to improve
the quality of phase retrieval, it is necessary to select an appropriate
coherence parameter and defocus distance to optimize the response of
the PTF.
At the end of this section, let us consider the question raised at the
beginning of this section. Teague [141] originally derived TIE under
the assumption of complete coherence, i.e., a monochromatic coherent
field. However, in practical implementations, TIE seems to work well
even when the illumination is not strictly coherent. In order to give a
reasonable explanation for this phenomenon, let us recall the defocused
coherent transfer function in the normalized coordinate [Eq. (76)]
̄ = 𝑃 (𝜌)
1
̄ 𝑒− 2 𝑗 𝑧̄ 𝜌̄
2
(269)
where
𝜌
𝜌̄ = (270)
𝑁𝐴𝑜𝑏𝑗 ∕𝜆
and
( √ )
𝑝𝑎𝑟𝑎𝑥𝑖𝑎𝑙
𝑧̄ = 2𝑘 1 − 1 − 𝑁𝐴2𝑜𝑏𝑗 𝑧 ≈ 𝑘𝑁𝐴2𝑜𝑏𝑗 𝑧 (271)
When analyzing the transfer function of a microscopic imaging system,
the normalized coordinate is often adopted to simplify the evaluation of
integrals. The scaling coefficients in Eqs. (73) and (74) are associated
with the transverse and axial extension of the 3D OTF, which will be
introduced in Section 8.
Substituting Eq. (269) into Eq. (250), we get the defocused WOTF
under the paraxial approximation
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 57. Transfer functions under different spatial coherence. (a) In-focus WOTFincoh , coherent parameters s ≥ 1; (b) defocused WOTFincoh , coherent parameters s ≥ 1;
(c) in-focus WOTFs , 𝑠 = 0.1, 0.4, 0.75 and 1.0 under different coherence parameters.

Fig. 58. The real part of WOTFs (ATF) for various coherent parameters and defocus distances (𝑁𝐴𝑜𝑏𝑗 =0.8, 𝜆 = 550𝑛𝑚, the spatial frequency coordinate is normalized
against the coherent resolution limit NAobj /𝜆). (a) 𝑠 = 0.1; (b) 𝑠 = 0.4; (c) 𝑠 = 0.75; (d) 𝑠 = 0.99.

⎧ √ ( )2 The defocused WOTF (under paraxial and weak defocusing approxima-

⎪∫ 𝑠−𝜌∕2
̄
2 𝑠2 − 𝜌2̄ + 𝑥 exp (𝑗 𝑧̄ 𝜌𝑥
̄ )𝑑𝑥 0 ≤ 𝜌̄ < 1 − 𝑠 tions) is obtained by substituting Eq. (274) into Eq. (250)
⎪ −(𝑠+𝜌∕2
̄ )
√
⎪ 𝑠−𝜌∕2 ( )2
̄ 𝑧̄ )=⎨∫ ̄ 2
𝑊 𝑂𝑇 𝐹𝑠 (𝜌, 2 𝑠2 − 𝜌2̄ + 𝑥 exp (𝑗 𝑧̄ 𝜌𝑥̄ )𝑑𝑥
⎪ −(1−𝑠 )∕2 𝜌̄ √
⎪ −(1−𝑆 2 )∕2 𝜌̄
( )2 √
𝜌̄ ⎧ 𝑠−𝜌∕2 ( )2 ( )
⎪+ ∫𝜌∕ 2 1 − 2 + 𝑥 exp (𝑗 𝑧̄ 𝜌𝑥̄ )𝑑𝑥 1 − 𝑠 ≤ 𝜌̄ ≤ 1 + 𝑠 ̄
⎩ ̄ 2−1 ⎪∫−(𝑠+𝜌∕2̄ )
2 𝑠2 − 𝜌2̄ + 𝑥 1 − 12 𝑗 𝑧̄ 𝜌̄2 𝑑𝑥 0 ≤ 𝜌̄ < 1 − 𝑠
⎪ √
(272) ⎪ ( )2 ( )
̄ 𝑧̄ )=⎨∫ 𝑠−𝜌∕2
𝑊 𝑂𝑇 𝐹𝑠 (𝜌, ̄
2 𝑠2 − 𝜌2̄ + 𝑥 1 − 12 𝑗 𝑧̄ 𝜌̄2 𝑑𝑥
⎪ −(1−𝑠 )∕2𝜌̄
2

Noted that this integral still cannot be evaluated analytically. But √ ( )2 ( )

⎪ −(1−𝑠2 )∕2𝜌̄
when the weak defocusing approximation is further introduced, i.e. ⎪+ ∫𝜌∕ 2 1 − 𝜌2̄ + 𝑥 1 − 12 𝑗 𝑧̄ 𝜌̄2 𝑑𝑥 1 − 𝑠 ≤ 𝜌̄ ≤ 1 + 𝑠
⎩ ̄ 2−1

(275)
Δ𝑧 ≪ 1∕𝜋𝜆𝜌2 (273)

the defocused coherent transfer function can be reduced to

Now the integral can be solved explicitly. The real part is completely
1
𝐻 (𝜌)
̄ ≈1 − 𝑗 𝑧̄ 𝜌̄2 (274) consistent with the in-focus ATF [Eq. (268)], and the imaginary part (i.e.,
2
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 59. The imaginary part of WOTFs (PTF) for various coherent parameters and defocus distances (𝑁 𝐴𝑜𝑏𝑗 =0.8, 𝜆 = 550𝑛𝑚, the spatial frequency coordinate is
normalized against the coherent resolution limit NAobj /𝜆). (a) 𝑠 = 0.1; (b) 𝑠 = 0.4; (c) 𝑠 = 0.75; (d) 𝑠 = 0.99.

PTF) is given by

⎧ 1 𝜋 𝑧̄ 𝜌̄2 𝑠2 0 ≤ 𝜌̄ ≤ 1 − 𝑠
⎪2 [ ( ) ( )]
⎪ 𝑧̄𝜌̄2 𝑠2 arccos 𝜌̄2 +𝑠2 −1 − arccos 𝜌̄2 −𝑠2 +1
⎪ 2 { 2𝜌𝑠
̄ 2𝜌̄
⎪ √ ( 2 2 )2 [ ( 2 2 ) ( 2 2 )]
⎪ 𝑧̄ 𝑠2 − 𝜌̄ +2𝑠𝜌̄ −1 𝑠2 arccos 𝜌̄ +2𝜌𝑠
𝑠 −1
− arccos 𝜌̄ −2𝑠𝜌̄ +1
̄ 𝑧̄ )=⎨+ 6𝜌̄
− Im 𝑊 𝑂𝑇 𝐹𝑠 (𝜌, ̄
⎪[( )2 2 ( )]
⎪ 1 − 𝑠2 − 𝜌̄ 1 + 𝜌̄2 + 7𝑠2
⎪ √ 2
( 2 2 )2 [(
⎪ 𝜌̄ −𝑠 +1 )2 2 ( )]
⎪+ 1 − 2𝜌̄
1 − 𝑠2 − 𝜌̄2 1 + 𝜌̄2 + 7𝑠2 1 − 𝑠 ≤ 𝜌̄ ≤ 1 + 𝑠
⎩
(276)

It can be found that when 0 ≤ 𝜌̄ < 1 − 𝑠, the PTF 12 𝜋 𝑧̄ 𝜌̄2 𝑠2 is just a

Laplacian in the Fourier domian. If we convert the normalized coordi-
nate back to the original coordinate, and neglect the constant scaling
factor (i.e., pupil area 𝜋s2 ), 12 𝑧̄ 𝜌̄2 → 𝜋𝜆z𝜌2 , it coincides with the TIE’s
PTF [Eq. (97)]. This suggests that under the paraxial approximation
and weak defocusing, TIE (without considering the partial coherence)
can still recover the phase information correctly within the low spa-
tial frequency range of 0 ≤ 𝜌̄ ≤ 1 − 𝑠. For spatial frequencies outside this
range, 1 − 𝑠 ≤ 𝜌̄ ≤ 1 + 𝑠, the form of PTF become much more compli-
cated. The amplitude of the PTF monotonically decays with the increase
of 𝜌, which can be observed in Fig. 59 (see the red curves corresponding
to Δ𝑧 = 0.5𝜇𝑚).
In order to illustrate the relationship between WOTF and TIE more
clearly, we show the ratio between the partially coherent weak defo-
cused PTF and TIE’s PTF in Fig. 60. It can be seen that the two PTFs
are in complete agreement within the range of 0 ≤ 𝜌̄ ≤ 1 − 𝑠, but be-
gin to deviate when 𝜌̄ becomes large. The divergence results from the
overestimate of phase contrast in TIE due to the ignorance of limited
aperture and partial coherence. Note that the trends of these curves are
very similar to that of in-focus partially coherent ATFs [Fig. 57(c)], but
Fig. 60. The ratio between imaginary part of the WOTF for a weakly defocused,
partially coherent bright-field microscope with coherent parameter s and the
inverse Laplacian filter of TIE.
their meanings are completely different. One major difference is that
when approaching incoherent illumination (s → 1), the PTF gradually
vanishes while the in-focus ATF still maintains a strong response.
7. Generalized TIE under partially coherent illuminations
No matter completely coherent fields, partially coherent fields, or
completely incoherent fields, the intensity of the optical field is well-
defined and directly accessible. Nevertheless, the “phase” is only well-
defined for coherent fields. In a strict sense, any physically realiz-
able light sources are not strictly coherent. Especially in adaptive op-
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

tics[144–148], X-ray diffraction imaging [161,162], neutron radiogra- imaging under partially coherent illuminations. They also proposed a
phy [164,165], and TEM [166,169,171–174], it is generally difficult to double exposure method to compensate for the phase errorn induced
obtain highly coherent light sources so that TIE becomes a simple and by the spatially partial coherence. Their conclusions are consistent with
effective alternative to interferometry. Even in the field of optical mi- the results of Streibl [142] and Gureyev et al. [352]. In 2015, Zuo et al.
croscopy, partially coherent illumination is beneficial to enhance the [208] derived the generalized TIE as an extension of the traditional TIE
imaging resolution, improve the image quality, and suppress the coher- for partially coherent fields in phase space. In this subsection, we intro-
ent noise. However, Teague’s TIE, as well as his derivation, assumes a duce two most representative extensions of TIE under partially coherent
monochromatic, coherent beam [141]. The 2D complex amplitude func- illuminations, i.e., partially coherent TIE based on MI/CSD and general-
tion is no longer sufficient to fully describe partially coherent fields, and ized TIE based on WDF.
4D MCF (or equivalently, CSD and WDF) should be used. Thus, Teague’s 7.1.1. Partially coherent TIE based on MI/CSD
TIE might encounter trouble when dealing with fields exhibiting non- From Subsection 6.1.4, we know that the propagation of CSD satisfies
negligible partial coherence. In this section, we analyze these issues and the following pair of Helmholtz equations [Eqs. (141) and (142)]
present the extensions of TIE for partially coherent fields. The challenges ( ) ( )
and opportunities brought by partially coherent illuminations for phase ∇21 𝑊𝜔 𝐱1 , 𝐱2 + 𝑘2 𝑊𝜔 𝐱1 , 𝐱2 = 0 (277)
retrieval and QPI is discussed. ( ) ( )
∇22 𝑊𝜔 𝐱1 , 𝐱2 + 𝑘2 𝑊𝜔 𝐱1 , 𝐱2 = 0 (278)
7.1. Extensions of TIE for partially coherent fields The propagation of MI also satisfies similar Helmholtz equations. In the
following, we adopt the CSD representation (note that the partially co-
The major obstacle in extending TIE for partially coherent fields herent TIE was originally derived by Streibl [142] based on MI in 1984.
arises from the fact that the partially coherent field does not have a Then in 2013, Petruccelli et al. [351] deduced the partially coherent
well-defined phase since the field experiences statistical fluctuations TIE in differential coordinate system based on CSD). Under the paraxial
over time. For partially coherent fields, the 2D complex amplitude is approximation, the Helmholtz equations reduce to paraxial wave equa-
insufficient to fully describe the properties of the field. They require the tions
( )
4D MCF or CSD of any two points in space, or equivalently being rep- ( ) 𝜕 𝑊𝜔 𝐱1 , 𝐱2
resented by WDF or AF in 4D phase space. In this case, the definition ∇21 𝑊𝜔 𝐱1 , 𝐱2 + 2𝑗𝑘 =0 (279)
𝜕𝑧
of phase should be clearly different from the one in a perfectly coher-
( )
ent field. As early as 1984, Streibl [142] used the 4D MI function to ( ) 𝜕 𝑊𝜔 𝐱1 , 𝐱2
reformulate TIE in order to make it explicitly account for the partial co- ∇22 𝑊𝜔 𝐱1 , 𝐱2 − 2𝑗𝑘 =0 (280)
𝜕𝑧
herence and demonstrated the possibility of using TIE for phase imaging In order to simplify Eqs. (279) and (280), the following differential co-
under partially coherent illumination. Although he did not give the QPI ordinate system is defined
results based on TIE (the numerical solution to TIE has not been pro-
𝐱 + 𝐱2 ′
posed at that time), this work is still extremely important because he 𝐱= 1 ; 𝐱 = 𝐱1 − 𝐱2 (281)
2
first pointed out that the phase retrieved by Teague’s TIE can still be
related to the phase of the object (not the field), provided that the in- or equivalently,
tensity distribution of the primary is axis-symmetric about the optical 𝐱′ 𝐱′
𝐱1 = 𝐱 + ;𝐱 = 𝐱 − (282)
axis. This pioneering work laid a preliminary theoretical foundation for 2 2 2
the subsequent extensive applications of TIE in optical microscopy. In In this case, it is not difficult to prove that ∇21 = ∇2𝐱 + ∇𝐱 ⋅ ∇𝐱 ′ +
1998, Paganin and Nugent [159] reinterpreted the “phase” of a partially 1 2
∇ , ∇22 = ∇2𝐱 − ∇𝐱 ⋅ ∇𝐱′ + 14 ∇2𝐱′ , and we can get = 2∇𝐱 ⋅ ∇𝐱′ .
∇21 − ∇22
4 𝐱′
coherent field as a scalar phase whose gradients become the normalized Then taking the difference between Eqs. (279) and (280) and simplify
transverse time-averaged Poynting vector of the field. The significance the result [142,351]
of this work lies in the fact that it endows “phase” with a completely new ( ) ( )
𝜕 𝐱′ 𝐱′ 𝑗 𝐱′ 𝐱′
and meaningful definition, providing a simple and reasonable physical 𝑊𝜔 𝐱 + , 𝐱 − = − ∇𝐱 ⋅ ∇𝐱 ′ 𝑊𝜔 𝐱 + , 𝐱 − (283)
basis for subsequent TIE phase retrieval under partially coherent illu- 𝜕𝑧 2 2 𝑘 2 2
minations. Unfortunately, this definition is only suitable for qualitative When x′ → 0, the LSH of Eq. (283) is just the PSD, thus the following
interpretation rather than quantitative analysis (because they did not transport equation can be obtained
( )
𝐱′ ||
give an explicit expression for the time-averaged Poynting vector of a 𝜕 𝑆 𝜔 (𝐱 ) 𝑗 𝐱′
partially coherent field). In 2004, Gureyev et al. [256] described an al- = − ∇𝐱 ⋅ ∇𝐱 ′ 𝑊𝜔 𝐱 + , 𝐱 − | (284)
𝜕𝑧 𝑘 2 2 ||𝐱′ =0
ternative interpretation with the generalized eikonal, based on the spec-
trum decomposition of a polychromatic field. He experimentally demon- Eq. (284) can be considered as the transport of spectrum equation for
strated that even under temporally partially coherent (polychromatic) polychromatic fields. For a quasi-monochromatic field, the spectral den-
illuminations, accurate phase retrieval can be obtained by TIE as long sity S𝜔 (x) just refers to the intensity I(x), and the partially coherent TIE
as the spectral distribution of the source is known. In 2016, Gureyev is obtained.
et al. [349] quantitatively analyzed the effect of partially coherent illu-
mination on TIE phase retrieval based on the CSD analysis. They derived 7.1.2. Generalized TIE in phase space
that the effects of both the spectral width and physical size of the source From Subsection 6.2.5, we know that the paraxial propagation of
(Schell model) can be interpreted as a convolution on the reconstructed WDF obeys the Liouville transport equation [141]
phase distribution, leading to phase blurring. For sources with a rela- 𝜕 𝑊𝜔 (𝐱, 𝐮)
+ 𝜆𝐮 ⋅ ∇𝐱 𝑊𝜔 (𝐱, 𝐮) = 0 (285)
tively high degree of coherence, the convolution effect is not obvious, 𝜕𝑧
so the phase distribution can still be accurately recovered. When the co- and the solution takes the form
herence of the light source is low, they also proposed to introduce one
𝑊𝜔 (𝐱, 𝐮, 𝑧) = 𝑊𝜔 (𝐱 − 𝜆𝑧𝐮, 𝐮, 0) (286)
additional sample-free measurement and use deconvolution algorithm
to compensate for the coherence-induced phase blur. In 2010, Zysk et al. where z is the propagation distance. In 2015, Zuo et al. [208] integrated
[350] explicitly considered the spatially partial coherence with use of on both sides of Eq. (285) over all spatial frequencies u, and combining
coherent mode decomposition, revealing that the phase recovered by the definition of PSD given by Eq. (171), derived that
TIE is a weighted average of the phases of all coherent modes. In 2014, 𝜕𝑆𝜔 (𝐱)
Petrucelli et al. [351] used the CSD representation to analyze TIE phase = −∇𝐱 ⋅ 𝜆𝐮𝑊𝜔 (𝐱, 𝐮)𝑑𝐮 (287)
𝜕𝑧 ∫
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Eq. (287) can be regarded as the transport of spectrum equation for

polychromatic fields. It relates the longitudinal evolution rate of the
PSD to the transverse divergence of the first frequency moment of the
WDF. The time-averaged intensity of a partially coherent beam coincides
with the integral of the PSD over all frequencies [256,350]. Hence, we
integrate Eq. (287) over 𝜔 to obtain the generalized TIE (GTIE) [208]
𝜕𝐼 (𝐱)
= −∇𝐱 ⋅ 𝜆𝐮𝑊𝜔 (𝐱, 𝐮)𝑑 𝐮𝑑 𝜔 (288)
𝜕𝑧 ∬
Only the paraxial approximation is employed in deriving Eq. (288), thus
it is general enough to cover various optical fields with arbitrary spatial
and temporal coherence.
Now let us consider a few important special cases of the GTIE: first,
when the field is quasi-monochromatic, i.e., the field consists of almost
a single frequency, the spectral density S𝜔 (x) is simply the intensity I(x).
In this case, the field can be regarded as almost completely temporally
coherent. Thus, the transport of spectrum equation reduces to the GTIE Fig. 61. Schematic of a simplistic view of coherent field and partially (spatially)
for spatial partially coherent fields: coherent field. (a) For coherent field, the surface of the constant phase is inter-
𝜕𝐼 (𝐱) preted as wavefronts with geometric light rays traveling normal to them. It is
= −𝜆∇𝐱 ⋅ 𝐮𝑊 (𝐱, 𝐮)𝑑𝐮. (289)
𝜕𝑧 ∫ fully described by 2D complex amplitude; (b) partially coherent field needs 4D
coherence function, like the Wigner distribution, to accurately characterize its
Note that although the temporal coherence of the illumination can be properties, like its propagation and diffraction. In addition, a partially coherent
simply incorporated by the integral over all optical frequencies as in field does not have a well-defined phase, but rather a statistical ensemble of
Eq. (288), it should be stressed that for dispersive samples, the inherent phases (spatial frequencies, propagation directions) at every position in space.
wavelength-dependent RI often complicates the accurate phase deter-
mination. For the remainder of this section, we will drop the spectral
dependence 𝜔 (assuming quasi-monochromaticity) for simplicity, but it can be regarded as a mono-component signal, and the first conditional
should be noted that for polychromatic fields, the WDF characterizes frequency moment of WDF (instantaneous frequency) is related to the
only a single spectral component of the whole field. transverse phase gradient of the complex field [303,308]
Similar to the traditional TIE, GTIE can also be derived from the en-
∫ 𝐮𝑊 (𝐱, 𝐮)𝑑𝐮 1
ergy conservation law of the electromagnetic field. In Subsection 6.2.5, = ∇ 𝜙(𝐱) (296)
[ ]𝑇 ∫ 𝑊 (𝐱, 𝐮)𝑑𝐮 2𝜋 𝐱
we construct the 3D vector field 𝑗𝐫 = 𝑗𝐱 , 𝑗𝑧 [303,323], which is also
known as the geometrical vector flux [Eqs. (206) and (207)] Substituting Eq. (296) into Eq. (295) leads to Teague’s TIE [141]
𝜕𝐼 (𝐱) 1 [ ]
𝑗 𝐱 (𝐮 ) = 𝜆 𝐮𝑊 (𝐱, 𝐮)𝑑𝐮 (290) = − ∇𝐱 ⋅ 𝐼 (𝐱)∇𝐱 𝜙(𝐱) (297)
∫ 𝜕𝑧 𝑘
√ As discussed earlier, the validity of Teague’s TIE is restricted to fully
1
𝑗𝑧 (𝐱) = 𝑘2 − 4𝜋 2 |𝐮|2 𝑊 (𝐱, 𝐮)𝑑𝐮 (291) coherent fields, while GTIE, which is a generalized version of TIE in
𝑘∫
phase space, explicitly considers the coherence modes so that it can be
The energy conservation in free space implies the geometrical vector
applied to a much wider range of optical- and electron-beams.
flux has zero divergence
One difficulty in extending GTIE to phase retrieval arises from
∇ ⋅ 𝑗𝐫 = 0 (292) the fact that the partially coherent field does not have a well-defined
“phase” since the field experiences statistical fluctuations over time
Then the following continuity equation can be obtained
(Fig. 61). It can be seen from Eq. (296) that GTIE relates the axial
𝜕 𝑗 𝑧 (𝐱 ) intensity derivative with the transverse divergence of the conditional
= −∇ ⋅ 𝑗𝐱 (𝐱) (293)
𝜕𝑧 frequency moment of WDF, while the phase gradient (defined as “gen-
√
eralized phase”) of partially coherent fields is automatically connected
Under the paraxial approximation, 𝑘2 − 4𝜋 2 |𝐮|2 ≈ 𝑘. In this case, the
with the conditional frequency moment of WDF (also known as the in-
axial energy flux reduces to the intensity, given by the space marginal
stantaneous frequency in the field of signal processing) [238]. It is not
of the WDF
difficult to find that the phase-space representation on the left side of
𝑗𝑧 (𝐱) ≈ 𝑊 (𝐱, 𝐮)𝑑𝐮 = 𝐼 (𝐱) (294) Eq. (296) is still valid for partially coherent fields, leading to a new
∫ meaningful and more general definition of “phase”. Here we refer the
Substituting Eq. (294) into Eq. (293), the GTIE for partially coherent new “phase” defined by Eq. (296) as the generalized phase of partially
fields is obtained [208] coherent fields to distinguish it from its coherent counterpart. It can
𝜕𝐼 (𝐱) be seen from Eq. (296) that the generalized phase is a scalar potential
= −∇ ⋅ 𝑗𝐱 (𝐱) = −𝜆∇𝐱 ⋅ 𝐮𝑊 (𝐱, 𝐮)𝑑𝐮 (295) whose gradient yields the conditional frequency moment of the WDF. It
𝜕𝑧 ∫
is clear from a distribution point of view that the quantity is the average
7.1.3. Generalized definition of “phase” for partially coherent fields spatial frequency at a particular location. Besides, the time-frequency
In the previous subsection, we drop the spectral dependence 𝜔 by as- joint description characteristic of WDF is very similar to the concept of
suming quasi-monochromaticity. However, quasi-monochromatic fields light in geometric optics. It supposes that light is a description of the
are still not necessarily deterministic due to the statistical fluctuations amplitude and propagation direction of light energy (Poynting vector).
over the spatial dimension. This randomness can be removed by fur- Thus, W(x, u) can be intuitively interpreted as the energy density of the
ther limiting the field to be completely spatially coherent as well. Then ray travelling through the point x and having a frequency (direction)
the field becomes deterministic√ and can be fully described by the 2D u. Eq. (286) exactly represents the geometric-optical behavior of a ray
complex amplitude 𝑈 (𝐱) = 𝐼(𝐱) exp [𝑗𝜙(𝐱)], where 𝜙(x) is the phase traveling through free space and Eq. (294) implies that the intensity at
of the completely (both temporally and spatially) coherent field. From a point is simply the sum of the energy spreading over all possible direc-
the space-frequency analysis perspective, the completely coherent field tions. Similarly, the frequency moment of WDF [Eq. (296)], represents
C. Zuo, J. Li and J. Sun et al.
the transversal ensemble/time-averaged flux vector (transversal time-
averaged Poynting vector) [159]. However, it should be noted that the
WDF is not a rigorous energy density function (radiance function) due
to its possibility of negativeness. But it is still very enlightening to use
this connection between physical optics and geometric optics. For the
cases when the diffraction effect is not obvious or even can be ignored, it
provides an accurate and meaningful physical explanation for the light
transport of partially coherent fields.
7.2. Phase retrieval under partially coherent illuminations
7.2.1. Phase retrieval based on GTIE
In this subsection, we explore the application of GTIE for phase re-
trieval under partially coherent illuminations. It should be emphasized
that the major concern in such scenario is the well-defined phase shift in-
troduced by the specimen, rather than the generalized phase of the par-
tially coherent field itself. This leads to the natural choice of treating the
contributions of the incident illumination and specimen separately by
considering the transmitted field as a product of the illumination func-
tion Uin (x) and the sample transmission function 𝑇 (𝐱) = 𝑎(𝐱) exp [𝑗𝜙(𝐱)],
where a(x) and 𝜙(x) are the amplitude and the phase of the specimen.
The CSD of the resultant field just leaving the object can be written
( ) ( ) ( ) ( )
as 𝑊𝑜𝑢𝑡 𝐱1 , 𝐱2 = 𝑇 𝐱1 𝑇 ∗ 𝐱2 𝑊𝑖𝑛 𝐱1 , 𝐱2 . Substituting it into the defi-
nition of the WDF and using the convolution theorem of the Fourier
transform, we can represent the overall WDF as a convolution of the
object transmittance WDF, WT (x, u), and the illumination WDF, Win (x,
u), over the spatial frequency variable 𝐮
𝑊𝑜𝑢𝑡 (𝐱, 𝐮) =𝑊𝑇 (𝐱, 𝐮)⊗ 𝑊𝑖𝑛 (𝐱, 𝐮)
𝐮 the two components (constant phase is unimportant in phase retrieval
( ) ( )
= 𝑊𝑇 𝐱, 𝐮′ 𝑊𝑖𝑛 𝐱, 𝐮 − 𝐮′ 𝑑 𝐮′ (298)
∫ the generalized phase of the illumination, and it only retrieves the gen-
Substituting Eq. (298) into the LHS of Eq. (296) and changing the order
of integration, it can then be derived that the generalized phase of the
transmitted field, 𝜙̂ 𝑜𝑢𝑡 (𝐱), should satisfy the following expression [208]
∫ 𝐮𝑊𝑜𝑢𝑡 (𝐱, 𝐮)𝑑𝐮 ∫ 𝐮𝑊𝑇 (𝐱, 𝐮)𝑑𝐮 ∫ 𝐮𝑊𝑖𝑛 (𝐱, 𝐮)𝑑𝐮
= + (299)
∫ 𝑊𝑜𝑢𝑡 (𝐱, 𝐮)𝑑𝐮 ∫ 𝑊𝑇 (𝐱, 𝐮)𝑑𝐮 ∫ 𝑊𝑖𝑛 (𝐱, 𝐮)𝑑𝐮
or equivalently,
[ ]
∇𝐱 𝜙̂ 𝑜𝑢𝑡 (𝐱) = ∇𝐱 𝜙̂ 𝑖𝑛 (𝐱) + 𝜙(𝐱) (300)
This representation shows the generalized phase accrues upon propaga-
tion through the object, behaving precisely as a conventionally defined
phase. The total generalized phase is the sum of the phase of the object
and the generalized phase of the incident illumination. In general, the
determination of the object phase requires two independent measure-
ments, performed with and without the presence of the specimen. The
sample-free measurement is used to characterize 𝜙̂ 𝑖𝑛 (𝐱) of the incident
beam and is subsequently subtracted from the total generalized phase
𝜙̂ 𝑜𝑢𝑡 (𝐱) to get the net phase introduced by the object only. However, if
the illumination is chosen judiciously to directly nullify 𝜙̂ 𝑖𝑛 (𝐱), then
𝐮𝑊𝑖𝑛 (𝐱, 𝐮)𝑑𝐮 = 0 (301)
∫ optical field, the redundancy of phase space becomes more obvious. As
Eq. (301) is called “zero-moment condition”. When this condition is satis-
fied, the total generalized phase 𝜙̂ 𝑜𝑢𝑡 (𝐱) directly gives 𝜙(x) and one sin-
gle measurement is sufficient to retrieval the object phase even though
the illumination is not fully coherent. Next, we will discuss two special
cases that satisfy Eq. (301). Firstly, for completely coherent illumina-
tion, Eq. (301) simply means the wavefront is on-axis plane wave, and
the waist of a Gaussian beam. For partially coherent spatially station-
ary illuminations [303] (discussed in Subsection 6.1.7), which will be
generally true for the experimental arrangements in optical microscopy
since they typically use Köhler illumination geometry, the spatially in-
coherent primary source (usually in the condenser aperture plane for an
optical microscope) featured by the intensity distribution S(x) and the
( ) ( )
positional CSD 𝑊 𝐱 + 𝐱′ ∕2, 𝐱 − 𝐱′ ∕2 = 𝑆 (𝐱)𝛿 𝐱′ is collimated by the
Optics and Lasers in Engineering 135 (2020) 106187
condenser (or simply propagated to the far field), producing the illumi-
nation WDF Win (x, u) just before the object plane [Eq. (200)]
𝑊𝑖𝑛 (𝐱, 𝐮) = 𝑆 (𝐮) (302)
Note that the above expression ignores the constant coordinate scal-
ing factor associated with the Fourier transform pair, which is trivial
when all computations are carried out in normalized units. Substitut-
ing Eq. (302) into Eq. (301) reveals that the primary source distribution
must be symmetric about the optical axis, which corresponds to the case
discussed by Streibl [142] and Petruccelli et al. [351].
Before proceeding further, we must emphasize that though GTIE is
derived in the joint space-spatial frequency domain using WDF, here we
do not intend directly to apply GTIE [Eq. (289)] or generalized phase
[Eq. (296)] for phase retrieval because WDF is difficult to measure di-
rectly, as discussed in Subsection 6.2.6. The main point to be conveyed
here is that phase retrieval can be realized by directly applying the orig-
inal Teague’s TIE for partially coherent fields, by adopting the new GTIE
and the generalized definition of “phase” that is valid for partially co-
herent fields. In other words, for completely coherent fields, if we take
the axial intensity derivative and then solve Teague’s TIE, we obtain
the phase of the field; for partially coherent fields, by following the
same procedure, we obtain the generalized phase of the partially co-
herent field instead. Since the conditional frequency moments of WDF
are additive [Eq. (299)], the gradient of the generalized phase also is
additive [Eq. (300)]. Hence, the generalized phase of the transmitted
field can be decomposed into the generalized phase of the incident illu-
mination plus the phase shift introduced by the specimen. This decom-
position is unique up to an additive constant that may float between
and can be neglected). GTIE knows nothing about the object phase and
eralized phase of the total beam passing through the object. However,
our objective is to determine the well-defined phase shift introduced by
the specimen rather than the “phase” of the illumination or the total
transmitted partially coherent field. To solve this problem, one needs
to either separate the two terms with two independent measurements
explicitly, as suggested in [351,352]. Alternatively, we can directly nul-
lify the “phase” of the illumination. As discussed above, for the coherent
imaging, illumination with flat wavefront is required; and for spatially
stationary illuminations, the primary source distribution must be sym-
metric about the optical axis. In fact, for the completely coherent case,
such kind of treatment has been habitually adopted in TIE literature. For
spatially stationary illuminations, it is also quite easy to realize in prac-
tice, e.g., the built-in Köhler illumination in a normal bright-field mi-
croscope (of course, the condenser aperture must be properly centered
on the optical axis, which is usually clearly explained in the microscope
operating manual).
7.2.2. Phase retrieval based on the definition of “generalized phase”
As mentioned above, the monochromatic coherent optical field can
be represented by a 2D complex amplitude function, so its 4D phase-
space representation is highly redundant. For a slowly varying coherent
shown in Eq. (190), the signal only fills a 2D section in phase space,
and the WDF is always strictly greater than 0, i.e., the non-negativity is
guaranteed. Therefore, for a slowly-varying coherent field, the diffrac-
tion effect can be approximately ignored, and the WDF is equivalent
to the energy density function (radiation brightness) of the optical field.
When spatial coordinates and spatial frequency coordinates are defined,
W(x, u) represented by Eq. (190) indicates that the light (energy flow)
passing through the point only propagates along one direction, which is
determined by the phase gradient (normal line). This property enables
phase measurement can be achieved by measuring the light direction,
which is just the basic idea of the Shake-Hartmann wavefront sensor
mentioned in Section 1 [40–42]. Fig. 62 shows a 1D diagram of WDF
and field distribution of a smooth coherent wavefront (spherical wave).
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 62. The WDF and light field of a smooth coherent wavefront. Phase is represented as the localized spatial frequency (instantaneous frequency) in the WDF
representation. Rays travel perpendicularly to the wavefront (phase normal). (a) Wavefront in real space; (b) WDF in phase space; (c) light field in position-angle
space.

Fig. 63. Principle of the Shack-Hartmann sensor and light-field camera. (a) For coherent field, the Shack-Hartmann sensor forms a focus spot array sensor signal; (b)
for a partially coherent field, the Shack-Hartmann sensor forms an extended source array sensor signa; (c) for incoherent imaging, the light-field camera produces a
2D sub-aperture image array.

It can be found that the direction of light propagation is perpendicular
to the wavefront (phase gradient direction). Under the paraxial approx-
imation, the relation between the spatial frequency in the WDF and the
light propagation angle in the light field can be simply summarized as
𝜽 ≈ 𝜆u, where 𝜽 represents the inclined angle of the light along the opti-
cal axis. Obviously, since the signal only fills a 2D section in phase space,
each microlens array behind the Hartmann wavefront sensor can only
approximately collect a highly localized 𝛿 function with different offset
[Fig. 62(a)], which reflects the high redundancy of the slowly varying
coherent optical field in the spatial frequency domain.
It is more complicated when the field is not strictly coherent. Gen-
erally, the phase-space WDF constitutes a rigorous and non-redundant
representation for partially coherent fields. From the perspective of ge-
ometric optics, for each point on the beam, light rays (energy flow) no
longer propagate in only one direction. Instead, they fan out to create a
2D distribution, i.e., a 2D sub-aperture image is created behind each mi-
crolens in the Hartmann wavefront sensor [Fig. 63(b)], which accounts
for the higher dimensionality of a partially coherent field. If the field
exhibits significant spatial incoherence, the phase-space negativity and
oscillations smooth out, and the WDF again approaches the radiance
or the light field. In the computer graphics community, the light-field
camera, as a counterpart of the Shack-Hartmann wavefront sensor, also
allows joint measurement of the spatial and directional distribution of
the incoherent light field [337] [Fig. 63(c)]. The “light field” usually
refers to as a collection of light, represented by L(x, 𝜽), where x is the
spatial location, and 𝜽 is the angular distribution of all light rays passing
through a given point in space. Under the geometric-optics approxima-
tion, the WDF is equivalent to radiance [300,301] or light field L(x, 𝜽)
[L(x, 𝜽) ≈ W(x, 𝜆u)] [335]. Since the intensity and angular distribution
of all rays are recorded by light-field imaging, ray-tracing technique
can be used to reconstruct synthetic photographs, estimate depth, and
change focus or viewing perspectives [353]. This is similar to the case
when the 4D coherence function of a partially coherent optical field is
retrieved, it allows to control the light propagation and manipulate the
optical field numerically in a computer. However, it significantly sacri-
fices spatial resolution as compared to conventional imaging systems.
As described in Subsection 6.2.6, it is difficult to measure the WDF
directly because of its high dimensionality, so it is not recommended
to directly use the definition of the generalized phase [Eq. (296)] for
phase retrieval. But it must be admitted that it is a very intuitive and
interesting approach for phase measurement. As is shown by Eq. (296),
the phase of the field (regardless of its state of coherence), which is
a scalar potential whose gradient yields the conditional frequency mo-
ment of the WDF, can be retrieved from TIE with a minimum of two
closely spaced intensity measurements. From Section 6, the 4D WDF
can be approximately acquired directly by using a microlens array (the
key component of a Shack-Hartmann sensor and light-field camera). Ap-
plying the geometric-optics approximation L(x, 𝜽) ≈ W(x, 𝜆u), we can
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 64. Images captured with a light-field microscope with different illumination NAs. (a) NAill = 0.05; (b) NAill = 0.15; (c) NAill = 0.2; (d) NAill = 0.25; scale bar
length is 50 𝜇m.

describe the phase in terms of the light field [Eq. (296)] can be changed
to
∫ 𝛉𝐿(𝐱, 𝛉)𝑑𝛉
= 𝑘−1 ∇𝜙(𝐱) (303)
∫ 𝐿(𝐱, 𝛉)𝑑𝛉
the quantity on the left side is just the centroid of the light field, i.e., the
average direction of light at one given position. Eq. (303) clearly reveals
that a standard TIE measurement can provide important (though not
complete) information of the light field, at least its angular marginal
and first angular moment. Conversely, it also tells us that the phase
gradient can be easily recovered from the 4D light field by a simple
centroid detection scheme. This is similar to the standard procedure in
the Shack-Hartmann method [40–42].
A simple experiment is presented here to verify the correctness of
Eq. (303). The experiment is based on a light field microscope built
upon a conventional microscope (Olympus BX41) with a microlens ar-
ray (pitch 150 𝜇m, ROC 10.518 mm) inserted in the intermediate image
Fig. 65. Phases reconstructed from the light fields with different illumination
plane just before the camera sensor]. The sample is a plano-convex mi-
NAs. (a) 𝑁 𝐴𝑐𝑜𝑛𝑑 = 0.05; (b) 𝑁 𝐴𝑐𝑜𝑛𝑑 = 0.15; (c) 𝑁 𝐴𝑐𝑜𝑛𝑑 = 0.2; (d) 𝑁 𝐴𝑐𝑜𝑛𝑑 = 0.25;
crolens array (pitch 100 𝜇m), which is imaged with a 20 × objective with scale bar length is 50𝜇m.
𝑁𝐴 = 0.4. Four light field images with the condenser NA from 0.05 to
0.25 are recorded, as shown in Fig. 64. From the enlarged images, we
can clearly see the intensity changing corresponding to each lenslet, understanding and quantifying the effect of the imaging system appears
from a focus spot array to a 2D sub-aperture image array. The centroid particularly important. Consider a practical imaging system with a finite
for each sub-image is calculated, followed by an integration to recon- aperture, the CSD in the image plane can be written as
struct the phase, as shown in Fig. 65. The results confirm that the phase ( ) ( ) ( )
𝑊𝑖𝑚𝑎𝑔𝑒 𝐱1 , 𝐱2 = 𝑊𝑜𝑢𝑡 𝐱1 , 𝐱2 ⊗ ℎ 𝐱1 , 𝐱2 (304)
can be extracted from the light field, even though the illumination is not 𝐱1 ,𝐱2
completely coherent. Note that when the condenser aperture is open up The mutual PSF h(x1 , x2 ) is defined as
to 𝑁𝐴𝑖𝑙𝑙 = 0.25, the sub-aperture images are too large to make them ( ) ( ) ( )
overlap, preventing reliable centroid detection. This results in artifacts ℎ 𝐱1 , 𝐱2 = ℎ 𝐱1 ℎ∗ 𝐱2 (305)
in the final retrieved phase image, as shown in Fig. 65(d). Neverthe- where h(x) is the coherent PSF of the imaging system. Based on the
less, due to limited number of microlens arrays and the size of CCD, the convolution property of WDF, the WDF in the image plane can be written
spatial resolution of the final reconstructed phase is quite low. These ex- as
perimental results prove the correctness of Eq. (303), showing that the ( )
𝐱′ 𝐱′ ( )
quantitative phase information can be directly extracted from the light 𝑊𝑖𝑚𝑎𝑔𝑒 (𝐱, 𝐮) = Γ𝑖𝑚𝑎𝑔𝑒 𝐱 + , 𝐱 − exp −𝑗2𝜋𝐮𝐱′ 𝑑 𝐱′
∫ 2 2
field (or WDF) based on the definition of “generalized phase”. However,
such an approach for phase retrieval is not recommended due to the low = 𝑊𝑜𝑢𝑡 (𝐱, 𝐮)⊗ 𝑊𝑝𝑠𝑓 (𝐱, 𝐮)
𝐱
spatial resolution and the experimental complexity as compared to the = 𝑊𝑇 (𝐱, 𝐮)⊗ 𝑊𝑖𝑛 (𝐱, 𝐮)⊗ 𝑊𝑝𝑠𝑓 (𝐱, 𝐮) (306)
full-resolution, defocus-based TIE techniques. 𝐮 𝐱

It is seen that the effect of the imaging system is equivalent to convolv-

7.2.3. Influence of the limited aperture of imaging system
Another critical assumption made by TIE is ideal imaging, which is
not fulfilled for a practical imaging system. In fact, what we measure
is the phase of the field in the image plane, not the phase of the object
itself, especially when the pupil of the imaging system is insufficient
to transmit all spatial frequencies of interest of the object. In this case,
ing the WDF of imaging PSF over the spatial variable x. More impor-
tantly, Wpsf (x, u) is zero when 𝐮 falls outside of the pupil (in most cases
the pupil function is equal to a circ-function, i.e., 𝑃 (𝐮) = 1, |u| ≤ uNA
and 𝑃 (𝐮) = 0, |u| > uNA ), which means all WDF components outside the
pupil will be dumped by the imaging system. From the definition of the
generalized function Eq. (296), TIE retrieves the conditional frequency
C. Zuo, J. Li and J. Sun et al.
Fig. 66. Numerical simulation on a pure phase sinusoidal grating. (a) Phase distribution of the simulated object; (b) 1D profile of the phase function; (c) Condenser
pupil function. (d) Objective pupil function; (e) The coherent PSF of the imaging system. All axes are expressed in the normalized units.
moment of the WDF as the phase gradient, so the reconstructed phase
in the image plane is not coincident with the true phase of the object
in general [expect for the case of perfect imaging, i.e., 𝑃 (𝐮) = 1, and the
Wpsf (x, u) reduces to 𝛿(x)]. Due to the bilinear nature of image forma-
tion in partially coherent imaging systems, such kind of phase discrep-
ancy is difficult to analyze or compensate directly. However, we con-
sider a slowly varying specimen under spatially stationary illumination,
Eq. (306) can be further simplified as [208]
[ ]
1
𝑊𝑖𝑚𝑎𝑔𝑒 (𝐱, 𝐮) ≈ 𝐼 (𝐱)𝑆 𝐮 − ∇𝐱 𝜙(𝐱) |𝑃 (𝐮)|2 (307)
2𝜋
Thus, the reconstructed gradient of the generalized phase in the im-
age plane is the frequency centroid of the overlapping area of the shifted
primary source and the pupil function. Without considering the effect of
the imaging system, the phase gradient is just the centroid of the shifted
primary source. As long as the source distribution is symmetric concern-
ing the optical axis, the phase can be accurately retrieved, regardless of
the source size (spatial coherence of the illumination). However, in a
practical imaging system, it is necessary to give higher importance to
the illumination coherence, because the size of the light source has a
significant influence on the imaging. Though decreasing the source size
does helps improve the phase retrieval accuracy (better linear transfer
for lower phase gradient), it will compromise the resolution limit. Fur-
thermore, a certain degree of illumination coherence is necessary. For
incoherent imaging (the source size is larger than the pupil), the real fre-
quency centroid corresponding to the object phase gradient can never
be correctly identified by TIE due to the pupil cut-off. For partially co-
herent imaging (the source size smaller than the pupil), the imaging
system induced phase distortion still exists but can be further compen-
sated, which is discussed in Subsection 7.2.4.
We present a series of numerical simulations to better illustrate the
above-mentioned theories. As shown in Fig. 66(a), the simulated object
is a pure phase sinusoidal grating with three different periods, 3 𝜇m,
1.5 𝜇m and 0.75 𝜇m respectively. To visualize the phase-space quan-
tities more conveniently, the sinusoidal grating is represented by a 1D
signal, as shown in Fig. 66(b). The object is illuminated by Köhler illumi-
nation with a circular condenser aperture NAill = 0.3, and the illumina-
tion intensity uniformly distributed over the aperture plane. The central
wavelength of the quasi-monochromatic illumination is 𝜆 = 550𝑛𝑚. The
object is imaged with an objective with NAobj = 0.7, thus the coher-
ent diffraction limit of the system is 0.7863 𝜇m. All computations are
Optics and Lasers in Engineering 135 (2020) 106187
performed in normalized units of 𝜆/NAobj for the space coordinate and
NAobj /𝜆 for the spatial frequency coordinate. As shown in Fig. 66(c) and
(d), in normalized coordinates, the radii of the S(u) and P(u) are s and
1, respectively, where s is the coherence parameter. The inverse Fourier
transform of the objective pupil gives the coherent PSF of the imaging
system, which is shown in Fig. 66(e).
Fig. 67 (a) shows the WDF 𝑊𝑇 (𝐱, 𝐮) of the specimen computed from
the object transmittance. According to Eq. (296), the phase derivative
obtained through TIE is equal to the first conditional frequency moment
of the WDF. To verify the accuracy of the phase retrieved by TIE, we
compared its derivative with the ideal phase derivative calculated from
the original phase function, shown in the normalized range from [-1,1]
[Fig. 67(b)]. The perfect match between the two curves indicates that
the phase can reliably be retrieved by TIE for the completely coherent
case. Next, we examine the case when the specimen is illuminated by
the partially coherent Köhler illumination, but assuming perfect imag-
ing conditions. According to Eq. (298), the WDF at the object plane
𝑊𝑜𝑢𝑡 (𝐱, 𝐮) can be calculated by convoluting the WDF of the object trans-
mittance WDF 𝑊𝑇 (𝐱, 𝐮) with the illumination WDF 𝑆 (𝐮), as illustrated
in Fig. 67(c). Compared with Fig. 67(a), the blurring of WDF along the
frequency dimension is clearly seen in Fig. 67(d). However, due to the
symmetry of the condenser aperture 𝑆 (𝐮), this blurring does not change
the frequency centroid of the object WDF. Thus, the phase derivative
can still be accurately retrieved, as verified by Fig. 67(e).
Next, the effect of the imaging system is considered. The WDF of
the imaging PSF, 𝑊𝑝𝑠𝑓 (𝐱, 𝐮), is shown in Fig. 67(f), and according to
Eq. (306) the image plane 𝑊𝑖𝑚𝑎𝑔𝑒 (𝐱, 𝐮) can be calculated by convoluting
𝑊𝑜𝑢𝑡 (𝐱, 𝐮) with 𝑊𝑝𝑠𝑓 (𝐱, 𝐮) along the 𝐱-direction, as shown in Fig. 67(g).
The imaging system removes all the WDF component falling outside of
the pupil, which in turn causes blurring in the retrieved phase deriva-
tive, as shown in Fig. 67(h). The imaging PSF greatly reduces the phase
contrast of the 0.75 𝜇m grating, but the phase structure for lower spatial
frequencies is less affected. It is instructive to further examine the effect
of the illumination coherence (i.e., the effect of gradually changing the
condenser aperture) when the imaging PSF is considered. This is also
consistent with our theoretical analysis.
Based on GTIE, Zuo et al. [208] quantitatively analyzed the image
formation and phase retrieval under partially coherent illuminations.
Through theoretical analysis and simulations, the following two conclu-
sions were obtained. First, when the primary light source is symmetric
about the optical axis, the phase of the object under partially coherent
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 67. Phase-space description of the effects of illumination and imaging system on TIE phase retrieval. See the text for details. (a) WDF of the specimen; (b)
retrieved phase derivative from (a); (c) WDF of the illumination; (d) WDF of the field in object plane; (e) retrieved phase derivative from (d); (f) WDF of the PSF of
the imaging system; (g) WDF of the field in image plane; (h) retrieved phase derivative from (g).

illuminations can be recovered by TIE through one-time measurement.
While when the primary source is not axisymmetric, the generalized
phase of the illumination can be measured by TIE (without the sample),
and then its influence can be compensated by subtracting the general-
ized phase of the illumination from the measured total phase (with the
sample). Secondly, although TIE does not impose any requirements on
illumination coherence, for a practical imaging system, a certain level
of spatial coherence is indispensable (an extreme case is when s ≥ 1, the
phase effect completely vanishes), and narrowing down the condenser
aperture a bit (𝑠 = 0.3 ∼ 0.5) is indeed conducive to accurate phase re-
trieval.
7.2.4. Phase gradient transfer function and coherent error compensation
In GTIE, the reconstructed generalized phase gradient just corre-
sponds to the centroid of the shifted primary source. Without consider-
ing the effect of lens aperture, reliable phase retrieval can be achieved as
C. Zuo, J. Li and J. Sun et al.
Fig. 68. Geometric interpretation of phase gradient recovered by TIE.
long as the primary source is symmetrical about the optical axis. How-
ever, for practical imaging systems, the frequency moment of WDF in
the image plane is deviated due to the cut-off effect of the objective
pupil, as shown in Fig. 68. The relation between the phase gradient re-
covered by TIE ∇𝐱 𝜙(̃ 𝐱) and the true phase gradient of the object ∇x 𝜙(x)
can be represented as
̃ 𝐱) ∫ 𝐮𝑊𝑖𝑚𝑎𝑔𝑒 (𝐱, 𝐮)𝑑𝐮
∇𝐱 𝜙(
= also different: PGTF is based on the slowly varying object approxima-
2𝜋 ∫ 𝑊𝑖𝑚𝑎𝑔𝑒 (𝐱, 𝐮)𝑑𝐮
[ ]
∫ 𝐮𝑆 𝐮 − ∇𝐱 𝜙(𝐱)∕2𝜋 |𝑃 (𝐮)|2 𝑑𝐮
= [ ] (308)
∫ 𝑆 𝐮 − ∇𝐱 𝜙(𝐱)∕2𝜋 |𝑃 (𝐮)|2 𝑑𝐮
To quantitatively characterize the attenuation effect of the imag-
ing system on different phase gradient (corresponding to different spa-
tial frequency) components, we can define a transfer function about
the phase gradient, so-called phase gradient transfer function (PGTF)
[343,354], which is the ratio between the measured phase gradient in
the image plane and the ideal phase gradient of the object
̃ 𝐱)
∇𝐱 𝜙(
𝑃 𝐺𝑇 𝐹 =
∇𝐱 𝜙(𝐱)
[ ]
1 ∫ 𝐮𝑆 𝐮 − ∇𝐱 𝜙(𝐱)∕2𝜋 |𝑃 (𝐮)|2 𝑑𝐮
= [ ] (309)
∇𝐱 𝜙(𝐱) ∫ 𝑆 𝐮 − ∇𝐱 𝜙(𝐱)∕2𝜋 |𝑃 (𝐮)|2 𝑑𝐮
Without considering the lens aperture effect (|𝑃 (𝐮)| = 1), the PGTF is
always 1 and the phase gradient can be recovered correctly. How-
ever, when the objective aperture is taken into account, the estimated
gradient will be smaller than the true value because of the source is
truncated by the pupil. The analytical expressions of PGTF for an ax-
isymmetric imaging system was derived by Sheppard et al. [354] un-
der different illumination conditions. Fig. 69 shows the PGTFs corre-
sponding to different coherence parameters. When 𝑠 = 1, it can be de-
duced from the geometric relationship between the light source and
the pupil function that the estimated gradient is always half of the true
phase gradient (PGTF is always 0.5). When s < 1, the low-gradient (low-
frequency) components corresponding to 0 ∼ s are consistent with the
real value, while the high-gradient (high-frequency) components are
underestimated. Meanwhile, it can be observed that there is a trade-
off between the resolution and accuracy of the phase reconstruction.
Large coherent parameters provides higher cut-off frequencies at the
expense of lower PGTF responses. In contrast, a small coherence param-
eter can provide unbiased estimates for low-gradient phase components,
Optics and Lasers in Engineering 135 (2020) 106187
Fig. 69. The phase gradient transfer function under different coherent param-
eters. Note that the abscissa represents the phase gradient value, but the coor-
dinate unit is the normalized spatial frequency.
but the imaging resolution is compromised. Note that the PGTF shares
some similarities with the ratio between the partially coherent WOTF
and TIE under weak defocusing [Eq. (60)]. However, they are essen-
tially quite different: PGTF is defined in the spatial domain and rep-
resents the transfer characteristics of the spatial gradient of the phase
function. In contrast, WOTF and PTF are defined in the frequency do-
main and represent the transfer characteristics of different frequency
components of the phase function. Their establishment conditions are
tion, while WOTF is based on the weak object approximation. When
𝑠 = 1, the WOTF becomes the incoherent OTF, and the phase informa-
tion completely vanishes, while the PGTF still maintains half of the
response.
As suggested by Zuo et al. [208,356], the PGTF can be used to com-
pensate the phase error induced by the spatial coherence of the illumi-
nation. A lookup table (LUT) can be built based on PGTF to calibrate
the phase gradient value recovered by TIE. The basic idea is similar to
deconvolution, but note that the compensation is performed on gradient
domain rather than Fourier domain. The entries of the LUT is the phase
gradient value recovered by solving TIE, and the output of the LUT is
the corresponding true phase gradient. The compensated phase gradi-
ent is then numerically integrated to obtain the phase that is free from
coherence-induced error [356].
7.3. Phase retrieval based on geometric-flow speckle tracking
7.3.1. Phase retrieval under coherent speckle illuminations
“Speckle” is a pattern generated when coherent light or partially co-
herent light illuminates an object consisting of randomly distributed
scatterers [357]. Although speckle should be usually avoided in laser
display [358] and coherent optical imaging [359], it has very impor-
tant applications in many fields, such as speckle imaging in astronomy
[360], ESPI in rough surface stress measurement [361], and dynamic
speckle imaging in biomedical research [362]. Recently, speckle illumi-
nation has been introduced into the field of phase retrieval and QPI.
In 2018, Paganin et al. [355] proposed a geometric-flow speckle track-
ing (GFST) method for X-ray phase-contrast imaging based on the con-
cept of geometric energy flow. As shown in Fig. 70, a paraxial forward
propagating beam illuminates a thin object and then traverses a short
distance z to the planar image sensor. IR (x) indicates the intensity of
the reference speckle, i.e., the image taken in the absence of the ob-
ject, where x is transverse coordinates in the plane perpendicular to the
optical axis z. The image in the presence of the sample is represented
C. Zuo, J. Li and J. Sun et al.
Fig. 70. Experimental setup of the geometric-flow speckle tracking method [355].
as IS (x). It is a resultant graph of distorted IR (x) due to the presence
of the measured sample, whose non-planar phase geometrically distorts
the reference speckles IR (x).
Assuming that the illumination is an ideal coherent plane wave, the
phase of the sample with different surface slopes will induce translation
of the speckle pattern, and the speckle displacement vector field ∇x d(x)
is proportional to the phase gradient ∇x 𝜙(x) [355]
𝑧 be obtained by solving TIE, and then the high-resolution phase distribu-
∇𝐱 𝑑(𝐱) = ∇𝐱 𝜙(𝐱) (310)
𝑘 tion can be reconstructed by phase integration. However, similar to the
where z is the distance between the sample and the detector. Therefore,
based on the relation between 𝑑(𝐱) and 𝜙(𝐱), the phase information of
the sample can be obtained by simply calculating the displacement vec-
tor field ∇𝐱 𝑑(𝐱) of the speckle pattern. In the GFST method, the flow is
defined as a conserved current associated with the deformation of the
reference speckles induced by the phase of the sample. By represent-
ing the transverse flow associated with deforming 𝐼S (𝐱) into 𝐼𝑅 (𝐱) as
the gradient of an auxiliary scalar potential function ∇𝐱 Λ = 𝐼𝑅 𝐷𝐱 , the
displacement field Dx can be obtained as
( { })
𝑗 ℱ [𝐼𝑅 (𝐱) − 𝐼𝑆 (𝐱)]
𝐷𝐱 (𝐱 )= ℱ −1 (𝑢, 𝑣) (311)
𝐼𝑅 (𝐱 ) 𝑢2 + 𝑣2
Assuming the field of the flow to be irrotational, the displacement field
can be described as the gradient of the scalar potential 𝑑𝐱 (𝐱)
( )
𝐷𝐱 (𝐱) ≡ 𝐷𝑥 (𝐱), 𝐷𝑦 (𝐱) ≈ ∇𝐱 𝑑(𝐱) (312)
Thus, the final formula for reconstructing the phase shift 𝜙(𝐱) based on
GFST is
( ) { }
𝐷𝐱 (𝐱) ⋅ 𝑘 𝑘 −1 ℱ [(𝐱0 + 𝑗 𝐲0 ) ⋅ 𝐷𝐱 (𝐱)]
𝜙(𝐱) = ∇−1
𝐱 = ℱ (313)
𝑧 𝑧 𝑗𝑢 − 𝑣
where x0 and y0 are unit vectors in x and y directions, respectively.
Compare with the FFT-based solver of TIE [Eq. (50)], it can be found
that GFST and TIE are essentially based on the same principle [356].
Both of them retrieve the phase information by solving the resultant
Poisson equations with the aid of scalar potentials, or so-called Teague’s
auxiliary function.
7.3.2. Phase retrieval based on partially coherent speckle illuminations
The idea of GFST is not just limited to X-ray diffraction imaging,
but also suitable for quantitative phase microscopy under partially co-
herent illuminations. Recently, Lu et al. [356] designed a QPI camera
with a weak diffuser (QPICWD) that is compatible with conventional
microscope platforms, as shown in Fig. 71. By combining the definition
of the generalized phase based on the energy flow (or current density),
the GFST was extended to QPI under partially coherent illuminations.
As described in Subsection 7.2.2, for each point on the partial coherent
fields, light rays (energy flow) no longer propagate in only one direction.
Optics and Lasers in Engineering 135 (2020) 106187
Instead, they fan out to create a 2D distribution. For partially coherent
fields, the distortion of the speckle field is a result of the statistical av-
erage of many rays passing through each spatial position and deflected
by the phase of the object. This is consistent with the generalized phase,
and the speckle displacement vector field ∇𝐱 𝑑(𝐱) is proportional to the
generalized phase gradient ∇𝐱 𝜙(𝐱), i.e., the first-order conditional mo-
ment of the WDF [356]. The high-resolution speckle distortion field can
case discussed in Subsection 7.2.4, the phase we measured is the gener-
alized phase of the “image” instead of the true phase of the object. Lu
et al. [356] analyzed the effect of illumination coherence on the resolu-
tion and accuracy of phase retrieval, revealing that the sample’s phase
can be reliably reconstructed when the coherence parameter is between
0.3 ∼ 0.5. This is consistent with the case of the uniform partially co-
herent Köhler illumination discussed in Subsection 7.2.3. Similarly, we
can also establish a LUT based on PGTF to compensate for the phase
blurring induced by the illumination spatial coherence [363].
7.4. Computational light-field imaging based on GTIE
For the TIE phase retrieval, it requires one in-focus image and one ad-
ditional defocused image to recover the phase information, which seems
plausible for the amount of information. Because the complex ampli-
tude is a 2D function, using the 2D defocused intensity information in
exchange for the 2D phase information appears “conservative” in the
amount of information. Thus, it is obviously highly redundant to rep-
resent a 2D coherent field in the 4D phase space. For a slowly varying
object, in Subsection 6.2.4, we mentioned that the redundancy of phase
space will become more obvious because the signal only occupies a 2D
section in phase space [Eq. (190)] [208]
[ ]
1
𝑊 (𝐱, 𝐮) = 𝐼(𝐱)𝛿 𝐮 − ∇𝜙(𝐱) (314)
2𝜋
The WDF now becomes nonnegative and takes on all the properties of
radiance. It clearly describes that the geometrical light ray at a single
position travels only along a single direction described by the phase
normal (coincides with the direction of the Poynting vector). It also tells
that the total amount of light carried by each ray is described by the
intensity of the field. This is an advantageous feature to allow phase
measurement simply by measuring the directions of rays, e.g., the Shack-
Hartmann sensor [41].
The situation becomes more complex when the optical field is not
strictly coherent. Generally, the phase-space WDF constitutes a rigorous
and non-redundant description for partially coherent fields. The knowl-
edge of amplitude and (generalized) phase are insufficient to determine
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 71. The schematic diagram of the quantitative phase microscope based on the weak diffuser camera [356].

the full field unambiguously. The measurement and retrieval of the com-
plete 4D coherence function (or equivalent, WDF and AF) are much
more complicated (see Subsection 6.2.6 for details). From the geometric-
optics perspective, for each point on the beam there exists many rays
with different directions; they fan out to make a 2D distribution, which
accounts for the higher dimensionality of the partially coherent field.
The light-field camera, as a counterpart of the Shack-Hartmann sensor
in the computer graphics community, allows joint measurement of the
spatial and directional distribution of light [337]. In the optics commu-
nity, light field is often known as “radiance” in radiometry [300,301].
As early as in 1968, Walther et al. [300] introduced the generalized
radiance, an equivalent of WDF, to lay a solid wave-optics foundation
for radiometry. In 2009, Zhang et al. [335] further clarified the equiva-
lence of light field and WDF under geometric-optics approximation, i.e.,
L(x, 𝜽) ≈ W(x, 𝜆u) [as shown in Figs. 62 (b) and (c)]. As discussed in
Subsection 6.2.6, the light-field camera directly captures the intensity
and angular distributions of all light rays, which can also be considered Fig. 72. Light field representation of a slowly varying object under spatially sta-
tionary illumination [208]. The sample exhibits angle-shift invariance: at each
as a coherence measurement technique. However, it requires elaborate
location, the direction of each incident ray shifts by the amount of object phase
optical setups (Fig. 49) and significantly sacrifices spatial resolution
gradient.
(traded for angular resolution) as compared to conventional imaging
techniques.
Although TIE cannot recover the complete 4D light field, it does pro- (slowly varying objects under spatially stationary illuminations), the 4D
vide important information about the light field. In interferometry, the light field is highly redundant (as shown in Fig. 72, the specimen can be
phase information is completely encoded into the fringe patterns, so the regarded as a “spreadless” or “angular shift-invariant system”: it does
coherent sources are required to produce observable fringes. However, not change the angular spread of the incident rays, which is fully deter-
TIE recovers the phase based on intensity propagation, and the intensity mined by the source intensity distribution. On the premise that the light
is always directly measurable, regardless of the degree of coherence of source distribution is known, the 4D light field can be fully characterized
the optical field. This property makes TIE much easier to implement. For by TIE [208].
partially coherent optical fields, the phase obtained by TIE is the gen- Under the geometric-optics approximation, L(x, 𝜽) ≈ W(x, 𝜆u), the
eralized phase [Eq. (296)] whose gradient is the first-order conditional physical picture behind Eq. (307) becomes quite clear. The angular dis-
frequency moment of WDF. Under geometric-optics approximation, the tribution of the light field just leaving the object is determined by the
WDF is equivalent to the light field L(x, 𝜽) ≈ W(x, 𝜆u), and the definition source intensity distribution shifted by the amount of the phase gradi-
of generalized phase then becomes [208] ent of the object. The imaging system only allows rays with the angles
within the pupil to pass and blocks the rest with larger angles. The in-
∫ 𝛉𝐿(𝐱, 𝛉)𝑑𝛉 tensity finally captured in the image plane is the sum of all rays passing
= 𝑘−1 ∇𝜙(𝐱) (315)
∫ 𝐿(𝐱, 𝛉)𝑑𝛉 through the imaging system, determined by the overlapping area of the
shifted primary source and the pupil function
The LHS of Eq. (315) is the centroid of the light field, i.e., the weighted
[ ]
average of the directions of light rays passing through a certain point in 1
𝐼𝑖𝑚𝑎𝑔𝑒 (𝐱) = 𝐼 (𝐱) 𝐮𝑆 𝐮 − ∇ 𝜙(𝐱) |𝑃 (𝐮)|2 𝑑𝐮 (316)
space. Based on Eq. (315), two conclusions can be drawn [208]. First, ∫ 2𝜋 𝐱
the 4D light field includes the 2D phase information, so that the phase Eq. (316) can be used to recreate 2D images of the sample from arbitrary
gradient can be easily recovered by localizing the centroid of each sub- perspectives: with the retrieved phase gradient ∇x 𝜙(x), one can simply
aperture image. This is similar to the standard procedure in the Shack- synthesize different views through Eq. (316) by shifting the position of
Hartmann method [40–42]. Second, TIE cannot recover the entire 4D the primary source artificially. Compared with the method presented
light field, but the conditional frequency moment (centroid) of the light in paper [364] (so-called pinhole renderings as in traditional light-field
field can be obtained. In addition, under certain simplified conditions imaging), which constructs the 4D light field first and then extracts its
C. Zuo, J. Li and J. Sun et al.
2D slices as perspective-shifted 2D images, Eq. (316) employs no empir-
ical assumptions and gives a more physically meaningful way for high-
resolution view synthesis with the effect of the imaging system taken
into account.
In fact, the first attempt to realize computational light-field imaging
based on TIE is the “light field moment imaging” (LMI) proposed by Orth
and Crozier [364] in 2013. They found that, with use of only a pair
of images exhibiting slight defocusing, the first angular moment of the
light field can be retrieved by solving a partially differential equation
(they did not notice that the equation is precisely TIE). With the mo-
ment retrieved, the perspective-shifted views of the original scene can
be synthesized. The “moment” in LMI just means that the method can
only recover the first angular moment, rather than the complete infor-
mation of the light field. To realize light-field imaging, Orth and Crozier
[364] assumed that the angular distribution of the light field is Gaussian
so that the missing light field data can be predicted. The empirical Gaus-
sian angular distribution assumption seems not physically founded, but
it does provides satisfactory parallax effect in real experiments. In 2014,
Zuo et al. [207] commented that the LMI is actually associated with TIE
at the geometric optics limit. Thus, all numerical solutions and axial in-
tensity derivative estimation algorithms in TIE can be directly applied
to the LMI, e.g., Liu et al. [365] used the multi-plane high-order finite-
difference method to improve the accuracy of axial intensity derivative
estimation, providing higher SNR and better visual effect for LMI.
7.5. Phase retrieval based on WOTF deconvolution
As demonstrated in previous subsections, although the definition of
phase for partially coherent fields is different from its coherent coun-
terpart, TIE can still be used to recover the phase of the object un-
der partially coherent illuminations. However, it should be noted that
GTIE is also based on weak defocusing approximation. However, in a
practical microscopic imaging system, a larger defocusing distance is
often preferred for better phase contrast, so the effect of defocusing dis-
tance on phase reconstruction cannot be simply ignored. In Section 6,
we quantitatively analyzed TIE’s PTFs under coherent and partially co-
herent illuminations (Fig. 59). Compare TIE’s PTF and the WOTF under
partially coherent illuminations, it is found that they only overlap at
low frequencies. As the degree of coherence decreases, the mismatch-
ing becomes more obvious. Therefore, when the phase is directly re-
constructed by TIE under partially coherent illuminations, the high-
frequency information will be excessively attenuated, resulting in the
loss of high-frequency details. To solve this problem, we can use weak
object approximation to achieve the linearization between intensity and
phase. This method is very similar to the coherent case, the only differ-
ence is that the coherent PTF in Subsection 4.3 should be replaced by
the partial coherent WOTF obtained in Subsection 6.3.5, and then fol-
lowed by a deconvolution corresponding to the WOTF to achieve more
accurate phase reconstruction. Another major advantage of this method
is that the range of linearization can be extended from weak defocusing
(near Fresnel zone) to an arbitrary defocus distance.
As discussed in Section 6.3.5, the real and imaginary parts of WOTF
corresponds to the ATF HA (u) and PTF HP (u), respectively. For an de-
focused axisymmetric imaging system, it can be found that HA (u) is an
even function of the defocus distance Δz, HP (u) is an odd function of
Δz. If we take two defocused images at opposite and equal defocuse
distances ± Δz along the optical axis, and calculate their normalized
difference in the Fourier domain
𝐼̂Δ𝑧 (𝐮) − 𝐼̂−Δ𝑧 (𝐮)
̂ 𝐮)
= −Im[𝑊 𝑂𝑇 𝐹 (𝐮)]𝜙( (317)
4𝐼̂0 (𝐮)
where 𝐼̂0 (𝐮)=𝑎20 𝑇 𝐶 𝐶 (0, 0) is the Fourier transform of the intensity distri-
bution at the in-focus plane. −Im[𝑊 𝑂𝑇 𝐹 (𝐮)] is the PTF of the partially
coherent imaging system, 𝜙( ̂ 𝐮) is the spectrum of the phase to be re-
trieved. It should be noted that the amplitude information is canceled
Optics and Lasers in Engineering 135 (2020) 106187
in the intensity difference, producing only phase contrast, and the phase
retrieval can be realized based on the WOTF deconvolution.
It should be noted that the response curve of the partially coherent
PTF HP (u) tends to be gradually oscillatory as the increase in defocus dis-
tance, leading to several zero-crossings (see Fig. 59). This phenomenon
also occurs under coherent conditions. The zero-crossings make the de-
convolution ill-conditioned, causing severe reconstruction artifacts. In
order to avoid the zero-crossings, we can reduce the defocused distance
to make Δz → 0 [256,257], thereby mitigating the oscillation of the PTF
at high frequencies. However, it can be seen from Fig. 59, when the
defocus distance is reduced, the phase contrast is also attenuated, espe-
cially for low spatial frequencies, making it difficult to obtain high-SNR
phase reconstructions [254,258]. Another solution is to capture more
intensity images at multiple defocus distances to synthesize the PTF,
just like the coherent case [200,256,259,260]. In this way, more spatial
frequencies can be more uniformly covered and zero-crossings can be
avoided in the synthesized PTF, so as to reduce the influence of noise
and improve the accuracy of phase reconstruction. For example, in 2014,
Jenkins et al. [366] proposed a multi-plane partially coherent phase re-
trieval approach termed multifilter phase imaging (MFPI), which is an
extension of the OSF method proposed by Zuo et al. [205] in the case of
coherent illumination. MFPI replaces the CTFs in OSF with the partially
coherent PTFs and recalculates the cut-off frequencies of the SGDFs with
different orders (Fig. 73). In 2015, Jenkins et al. [367] proposed another
multi-plane partially coherent phase retrieval method based on WOTF
deconvolution. The weighting functions for the intensities at different
defocus distances are optimized based on the linear least-square method.
It is worth mentioning that the SGDF and the least-square WOTF decon-
volution methods share similar ideas. SGDF is essentially a least-square
approach in axial derivative estimation (SGDF is the convolution form of
the least-square fitting) [205]. The phase reconstructed from each SGDF
is calculated based on all the intensity measurements. In the least-square
WOTF deconvolution method, one phase reconstruction is based on the
intensity difference between a pair of symmetrically defocused images,
so it does not make full use of all intensity measurements. Recently, Bao
et al. [368] compares and contrasts these two approaches in detail, re-
vealing that the SGDF approach is more accurate than the least-square
WOTF deconvolution method but is also slower in computation.
7.6. Resolution enhancement based on coherence engineering
In Subsection 6.3.5, we learned that for a coherent imaging system,
the defocus PTF has a strong response, i.e., the captured intensity image
can provide a relatively high phase contrast. Nevertheless, the imaging
resolution is limited to the coherent diffraction limit. Partial coherent
imaging extends the maximum achievable imaging resolution beyond
the coherent diffraction limit. The resolution limit of partially coherent
imaging is determined by the sum of the NA of the objective lens and NA
of the illumination. As the coherence parameter increases, the theoreti-
cal imaging resolution also improves. However, it also causes a signifi-
cant reduction of the response of the PTF, exacerbating the ill-posedness
of WOTF deconvolution. Therefore, for a conventional bright-field mi-
croscope with a circular condenser diaphragm, in order to achieve a
compromise between the imaging resolution and phase contrast, the co-
herence parameter should be generally set between 0.3 and 0.5 [208].
Although multi-plane approaches utilizing intensity measurements at
both small and large defoci allow the response of PTF to be optimized
over a wider range of spatial frequency, the noise-to-resolution tradeoff
is still not fundamentally solved. As is predicted by the WOTF analysis
(Fig. 59), the phase contrast progressively vanishes as the illumination
NA approaches the objective NA, suggesting the phase information can
hardly be transferred into intensity via defocusing when illumination
NA is large. This poses a fundamental obstacle to improving the res-
olution of TIE phase imaging up to twice of coherent diffraction limit
(diffraction limit of incoherent imaging). It should also be mentioned
that although synthetic aperture techniques via oblique [369] or struc-
C. Zuo, J. Li and J. Sun et al.
Fig. 73. The block diagram of quantitative phase recovery method based on OFS with multi-plane TIE under partially coherent illuminations.
tured illumination [370,371] have been demonstrated to be possible
solutions to enhance the phase imaging resolution of TIE, most of them
require relatively complicated optical systems which are not typically
available to most bio/pathologists, prohibiting their widespread use in
biological and medical science.
In the TIE phase retrieval, the distribution of the PTF is determined
by the defocus distance and the coherence parameter. Nevertheless,
there is another very important adjustable parameter that is not consid-
ered in our previous discussions: the shape of the illumination aperture.
The condenser aperture of conventional microscopes is roughly circular
in shape, and the spatial coherence of the illumination (coherence pa-
rameter s) can be tuned by changing the radius of condenser diaphragm.
However, the shape of the condenser aperture is not just limited to cir-
cular. Zernike’s phase-contrast microscope is a very good example: a
specially designed annular diaphragm, which is matched in diameter
and optically conjugate to an internal phase plate residing in the ob-
jective rear focal plane, is placed in the condenser front focal plane to
modulate the illumination aperture. Recently, it has been found that co-
herence engineering by changing the illumination aperture is generally
more effective than changing the defocus distance for optimizing the
WOTF [209,210,342,372]. In 2017, Zuo et al. [209] suggests to replace
the conventional circular illumination aperture with an annular one
{ lumination (91% for Δs = 0.01, 82% for Δs = 0.1), but the response
1 𝑠1 ≤ |𝜌|̄ ≤ 𝑠2
𝑆 (𝜌)
̄ = (318)
0 |𝜌|̄ < 𝑠1 , |𝜌|
̄ > 𝑠2
where s1 and s2 are the normalized inner and outer diameters of the
annular aperture, respectively. The corresponding PTFs (the imaginary
part of the WOTF) with different combinations of s1 and s2 can be
calculated based on Eq. (250), which are shown in Figs. 74 and 75.
Fig. 74 shows the PTFs of annular illuminations with fixed annulus
width but different inner and outer diameters. It can be found that
when the annular illumination has a smaller inner diameter, the re-
sponse amplitude of the PTF is large, but the cut-off frequency of the
PTF is still limited to near the coherent diffraction limit. However, as
the inner and outer diameters of the annular illumination increase, the
cut-off frequency of PTF is extended to the incoherent diffraction limit,
and the overall curve is gradually moved from the first quadrant to the
forth quadrant. This result indicates that an annular illumination source
with a maximum outer diameter and inner diameter not only provides
non-coherent diffraction-limited imaging resolution but also provides a
strong amplitude response in the cut-off frequency passband. In Fig. 75,
we further demonstrate the effect of varying the thickness of the an-
nulus by fixing the NA of the outer circles to be 1 (s2 = 1) and only
changing the thickness of the annulus (𝑠1 = Δ𝑠 from 0 to 1). As might be
Optics and Lasers in Engineering 135 (2020) 106187
expected, the phase contrast is reduced as the annulus width increases.
When Δs → 1, the phase contrast finally goes to zero, which is just the
incoherent case of the circular illumination. It is also shown that the cut-
off frequency of the WOTF is reduced from 2 to 2 − Δ𝑠 with the increase
of the annulus width. From the results shown in Figs. 74 and 75, it can
be concluded that we should choose the diameter of the annulus to be
equal to that of the objective pupil, and make its thickness as small as
possible to optimize both phase contrast and imaging resolution.
In Fig. 76, we compare the magnitudes of the PTFs of the annu-
lar illuminations (Δs = 0.01 and Δs = 0.1) and circular illuminations
(𝑠 = 0.1, 0.75, 0.99) under weak defocusing (Δ𝑧 = 0.5𝜇𝑚). It can be ob-
served that the annular illumination provides strong responses inside
the passband as well as a spatial frequency cut-off near the incoher-
ent diffraction limit (1.99NAobj for Δs = 0.01, 1.9NAobj for Δs = 0.1).
Compared with a conventional microscope with a wide-open circular
aperture (𝑠 = 0.75), the total phase contrast (the area enclosed by the
PTF curve and the frequency-axis) provided by the annular illumination
is more than doubled (2.35 times for Δs = 0.01, 2.11 times for Δs =
0.1). Not only the frequency coverage is extended, but the response in
both low and high spatial frequencies is significantly enhanced. The to-
tal phase contrast provided by the annular illumination is comparable
to that of a conventional microscope with nearly coherent (s = 0.1) il-
is much smoother and more extensive. The spatial frequency cut-off
is almost doubled, and the phase contrast of low spatial frequencies
is significantly increased. These results demonstrate that replacing the
conventional circular aperture with an annular one provides a conve-
nient way to optimize the WOTF for achieving a broadband frequency
coverage and enhanced response in both low- and high-frequencies.
Moreover, the resulting PTF contains no deep dips and zero-crossings
in its pass-band, which removes the ill-posedness of the WOTF inver-
sion. It is expected to achieve high-quality phase reconstruction and
overcome the noise-resolution tradeoff in TIE under partially coherent
illuminations.
In Fig. 77, we compare the phase retrieval results of annular illu-
mination TIE and circular illumination TIE based on simulations. The
Siemens star image is used as an example phase object [shown in
Fig. 77(a)] which is defined on a grid with 256 × 256 pixels with a pixel
size of 0.13𝜇m × 0.13𝜇m. The wavelength of the illumination is 550nm,
and the NAobj is 0.80. For such an imaging configuration, the best phase
imaging resolution can be achieved is 344nm (𝜆/2NAobj ), which is also
shown in Fig. 77(a). To simulate the noise effect, each defocused im-
age is corrupted by Gaussian noise with a standard deviation of 0.01.
Fig. 77(b) compares the defocused images, and the phase retrieval re-
C. Zuo, J. Li and J. Sun et al.
Fig. 74. Phase transfer function of annular illumination apertures with fixed annulus width but different inner and outer diameters at different defocus distances
(𝑁𝐴𝑜𝑏𝑗 = 0.8, 𝜆 = 550𝑛𝑚, the spatial frequency coordinate is normalized against the coherent resolution limit NAobj /𝜆). (a) 𝑠 = 0.1; (b) 𝑠 = 0.4; (c) 𝑠 = 0.75; (d) 𝑠 = 1.0.
sults of different illumination settings for a small defocus distance (Δz
= 0.5𝜇m). The metric used to measure the accuracy of phase retrieval
is given by the root mean square error (RMSE), which quantifies the
overall difference between the true phase and the retrieved phase. For
the case of circular illumination, the overall phase contrast reduces with
the increase in coherent parameter s, which is in coincidence with the
WOTF analysis (Fig. 59). The poor response at low-spatial frequencies
leads to cloud-like artifacts superimposed on the reconstructed phases.
Besides, the phase imaging resolution is improved by opening up the
condenser diaphragm (increasing the coherence parameter s). However,
for the case of nearly matched illumination (s = 0.99), the washout in
phase contrast prevents any recognizable phase information to be re-
constructed, leading to significant artifacts and a very large RMSE. The
phase contrast, especially for low-frequency components can be signif-
icantly enhanced by using the annular illumination. The Siemens star
appears dark in the defocused image, demonstrating the negative phase
contrast as predicted by the theory. The strong phase contrast is finally
converted to the quantitative phase images by WOTF inversion, result-
ing in high-quality reconstructions with a uniform background and im-
proved resolution. The RMSE values for the AI-TIE are comparable with
the conventional two-distance TIE approaches and significantly lower
than the case when only a single defocus distance is used. Besides, the
theoretical resolution for AI-TIE is improved to 1.9NAobj for Δs = 0.1
and 1.99NAobj for Δs = 0.01, which approaches to the incoherent limit
[RHS of Fig. 77(a)]. The above simulation results suggest that use of an-
nular illumination matching the objective NA allows for high-quality,
low-noise phase reconstruction with a lateral resolution close to the in-
coherent diffraction limit, providing significant resolution improvement
over its circular alternatives.
Optics and Lasers in Engineering 135 (2020) 106187
The above theoretical analysis and experimental results suggest that
reshaping of illumination source provides new possibilities to enhance
the imaging resolution and improve low-frequency performance of TIE
imaging. However, the choice of annular aperture was empirically de-
signed based on intuitive criteria related to the shape of WOTF. It is
still unclear whether the annular illumination is the best choice for TIE
phase imaging. Due to the complicated form of WOTF and high degree
of freedom of the aperture function, solving for an optimum source pat-
tern analytically seems quite challenging. In 2018, Li et al. [211] de-
veloped a numerical scheme to optimize the illumination pattern based
on a combined quantitative criterion for evaluating the “goodness” of
an aperture. In order to make the size of the solution searching space
tractable, only binary-coded axis-symmetric illumination patterns are
considered. Note that the axis-symmetric illumination satisfies the “zero-
moment condition” [Eq. (301)], which is a precondition for unbiased TIE
phase retrieval under partially coherent illuminations [208]. As shown
in Fig. 78, the light source is divided radially, and the pupil of the circu-
lar incoherent lighting source is separated into many concentric annuli
with equal spacing. The concentric annuli are represented by a 12-bit
binary number, and all the illumination source patterns can be indexed
by a decimal value of the corresponding binary number.
The PTFs corresponding to different source patterns are calculated.
Fig. 79 shows five typical illumination patterns and their corresponding
PTFs. The properties of PTF directly determines phase imaging perfor-
mance, i.e., the cut-off frequency determines the imaging resolution, the
number of zero-crossings indicates the ill-posedness of WOTF deconvo-
lution, and the area enclosed by the PTF and the coordinate axis rep-
resents the total phase contrast (the information-bearing portion of the
image). Therefore, the PTFs are evaluated based on a combined quanti-
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 75. Phase transfer function of different defocus distances under annular illumination with maximum annular outer diameter but different inner (𝑁𝐴𝑜𝑏𝑗 = 0.8,
𝜆 = 550𝑛𝑚, the spatial frequency coordinate is normalized against the coherent resolution limit NAobj /𝜆). (a) Δ𝑠 = 0.1; (b) Δ𝑠 = 0.4; (c) Δ𝑠 = 0.75; (d) Δ𝑠 = 0.99.

Fig. 76. Magnitude comparison of the phase transfer functions of the annular illuminations (𝑠 = 0.1 and 𝑠 = 0.01) and circular illuminations (𝑠 = 0.1, 0.75, 0.99) when
the defocus distance is 0.5 𝜇m. (a) Phase transfer function curves of different illuminations; (b) enlarged curve of low frequency component (blue boxed region); (c)
enlarged curve of high frequency component (red boxed region).

tative criterion. First, the cut-off frequency of the PTF must reach twice
the NA of the objective, which means the final imaging resolution is
extended to the incoherent diffraction limit. Then, the PTF does not in-
tersect with the coordinate axis within the cut-off frequency, i.e., there is
no zero-crossing in the PTF. Finally, the area enclosed by the PTF should
be maximized to guarantee the best SNR. By comparing the PTFs based
on the combined criterion, the optimal illumination pattern was identi-
fied, which is just a thin annulus matching the NA of the objective, as
we mentioned above.
The optimality of the annular illumination is experimentally verified
through experiments, as shown in Fig. 80. In order to control the illumi-
nation patterns flexibly, the light source of a conventional bright-field
microscope is replaced by a programmable LED array. Fig. 80 shows
the illumination patterns, corresponding PTFs, axial intensity deriva-
tives and Fourier spectra, and final reconstructed phases under differ-
ent illumination patterns. It can be seen that the PTF corresponding to
the matched annular illumination provides the highest imaging resolu-
tion and strongest phase contrast, especially for low-frequency compo-
nents. Thus, the corresponding reconstructed phase has the best con-
trast, resolution, and SNR. Finally, it is worth mentioning that the an-
nular illumination scheme and the multi-plane TIE approaches are not
contradictory in principle, and they can be combined to gain comple-
mentary advantages. For example, the OSF method and least-square de-
convolution method can be extended to annular illumination by sim-
ply replacing the coherent CTFs to WOTFs under annular illuminations
[373].
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 77. Comparison between annular illumination TIE and circular illumination TIE. (a) The raw Siemens star image and the corresponding best diffraction limited
image can be achieved based on the simulation parameter; (b) Comparison of over-defocus images and reconstruction results of different illumination settings for a
small defocus distance (Δz = 0.5 𝜇m).

Fig. 78. Illumination source encoding scheme. The circular incoherent source is divided into 12 concentric annuli, which can be indexed by a 12-bit binary number.
All possible illumination source patterns can be uniquely indexed by a decimal value of the corresponding binary number.

8. 3D phase imaging under partially coherent illuminations
When we talk about QPI techniques, it is usually assumed that the
measured samples are 2D (thin) objects, which can be represented as
a 2D complex transmittance function composed of absorption compo-
nent and phase component. The complex amplitude distribution of the
transmitted optical field is the product of the complex amplitude of the
incident optical field and that of the object. However, the phase delay in-
duced by the object is actually the axial projection (accumulation) of its
3D RI distribution onto a 2D plane (commonly known as 2.5D imaging),
which is an integral along the light propagation direction instead of real
3D information [374–377]. This problem can be effectively overcome by
optical diffraction tomography (ODT) techniques [375,376,378,379],
which enables high-resolution real 3D (lateral + axial) imaging of the
3D sample by recovering its 3D RI distribution. Conventionally, this
method needs to combine the phase measurement technique (e.g., digi-
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

a sphere in 3D Fourier space with the radius of 𝜆𝑛 , the so-called “Ewald

sphere”. In particular, if the lateral spatial frequency component (ux ,
uy ) of the optical field is determined, the z component of the spatial
frequency vector can be represented by
√
( )2
𝑛
𝑢𝑧 = ± − 𝑢2𝑥 − 𝑢2𝑦 (322)
𝜆
According to the angular spectrum theory introduced in
Subsection 2.2.2, once the complex amplitude of the coherent field
U(x, y, 0) in a certain plane is known (without loss of generality, we
consider 𝑧 = 0), the complex field U(x, y, z) at a arbitrary distance of z
can be determined
√
2 2
𝑈̂ 2𝐷 (𝑢𝑥 , 𝑢𝑦 , 𝑧) = 𝑈̂ 2𝐷 (𝑢𝑥 , 𝑢𝑦 , 0)𝑒𝑗𝑘𝑧 1−(𝜆𝑢𝑥 ) −(𝜆𝑢𝑦 )
√
Fig. 79. The phase transfer functions and corresponding 1D cross-sections for ( )2
1
𝑗2𝜋𝑧 −𝑢2𝑥 −𝑢2𝑦
5 different illumination patterns. = 𝑈̂ 2𝐷 (𝑢𝑥 , 𝑢𝑦 , 0)𝑒 𝜆
(323)

Taking Fourier transform on z of both sides of Eq. (323) yields

tal holography or phase retrieval) with the CT technique: first obtains a √
( )2
1
set of quantitative phase projections by rotating the object or changing 𝑗2𝜋𝑧 −𝑢2𝑥 −𝑢2𝑦 −𝑗2𝜋𝑧𝑢
𝑐 𝑈̂ 3𝐷 (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ) = 𝑈̂ 2𝐷 (𝑢𝑥 , 𝑢𝑦 , 0)𝑒 𝜆
𝑒 𝑧 𝑑𝑧
the illumination directions, and then reconstructs the 3D RI distribution ∫
( √ )
based on the CT principle. In recent years, “transport of intensity diffrac- ( )2
1
−𝑗2𝜋𝑧 𝑢𝑧 − −𝑢2𝑥 −𝑢2𝑦
tion tomography” (TIDT) has gradually emerged as a new diffraction 𝜆
= 𝑈̂ 2𝐷 (𝑢𝑥 , 𝑢𝑦 , 0)𝑒 𝑑𝑧
tomography technique based on the principle of non-interferometric, ∫
√
intensity only measurement inspired by TIE [214,380,381]. Compared ⎛ ( )2 ⎞
1
with traditional ODT techniques, this method does not require coher- = 𝑈̂ 2𝐷 (𝑢𝑥 , 𝑢𝑦 , 0)𝛿 ⎜𝑢𝑧 − − 𝑢2𝑥 − 𝑢2𝑦 ⎟ (324)
⎜ 𝜆 ⎟
ent illumination and interferometric measurement. It only needs to cap- ⎝ ⎠
ture the through-focus intensity images across the object, and then uses
the image deconvolution algorithm to directly retrieve the 3D RI dis- Eq. (324) shows that the coherent optical field is highly redundant
in 3D Fourier space,
( and non-zero values
) can only be taken on the
tribution of the object. Thus, it can effectively bypass the difficulties √( )
2
associated with the traditional interferometric diffraction tomography Ewald sphere. 𝛿 𝑢𝑧 − 1
− 𝑢𝑥 − 𝑢𝑦 denotes the projection of a 2D
2 2
𝜆
techniques, such as interferometric measurements, object rotation, and
beam scanning. This section is devoted to the study of diffraction to- Fourier spectrum onto the Ewald hemisphere surface. The direction of
mography and TIDT. the hemisphere depends on the light propagation direction with respect
to z-axis. Therefore, the 2D Fourier spectrum in 𝑧 = 0 plane actually con-
8.1. 3D Fourier spectrum and Ewald sphere for coherent fields tains all the information about the 3D Fourier spectrum of 3D optical
field
In Subsection 2.2, we have learned that any 2D deterministic coher- | √
| ( )2
ent fields can be decomposed into a superposition of plane waves with | 1
𝑈̂ 2𝐷 (𝑢𝑥 , 𝑢𝑦 , 0) = 𝑈̂ 3𝐷 (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 )|𝑢𝑧 = − 𝑢2𝑥 − 𝑢2𝑦 (325)
different propagation directions, which is known as the angular spec- | 𝜆
|
|
trum representation. Generalizing this rule into 3D space, the complex
amplitude distribution of the stationary coherent optical field is assumed Similarly, the 2D Fourier spectrum in the defocused plane Δz can also
to be U(x, y, z), and its Fourier spectrum in 3D space can be connected be associated with the 3D Fourier spectrum
by the following 3D Fourier transform √ √
( )2
⎛ ( )2 ⎞ 1
1 𝑗2𝜋Δ𝑧 −𝑢2𝑥 −𝑢2𝑦
𝑈̂ 2𝐷 (𝑢𝑥 , 𝑢𝑦 , Δ𝑧) = 𝑈̂ 3𝐷 ⎜𝑢𝑥 , 𝑢𝑦 , − 𝑢𝑥 − 𝑢𝑦 ⎟𝑒
2 2 𝜆
𝑈 (𝑥, 𝑦, 𝑧) = 𝑈̂ 3𝐷 (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 )𝑒𝑗2𝜋(𝑢𝑥 𝑥+𝑢𝑦 𝑦+𝑢𝑧 𝑧) 𝑑 𝑢𝑥 𝑑 𝑢𝑦 𝑑 𝑢𝑧 (319) ⎜ 𝜆 ⎟
∭ ⎝ ⎠
where the exponential primitive 𝑒𝑗2𝜋(𝑢𝑥 𝑥+𝑢𝑦 𝑦+𝑢𝑧 𝑧) represents a plane wave = 𝑈̂ 3𝐷 (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 )𝑒𝑗2𝜋𝑢𝑧 Δ𝑧 (326)
exp (jk · r) propagating in 3D space. Considering an instantaneous shot
of a plane wave propagating in 3D space, wave peaks and troughs can be 8.2. 3D coherent transfer function and generalized aperture
observed, as shown in Fig. 81. The distance from the peak to the trough
corresponds to the wavelength of the light wave (when the light propa- In Section 4, we have learned that the coherent imaging system is lin-
gates in non vacuum, it is the ratio of light wave to RI of medium). The ear about the complex amplitude, which is completely determined by its
frequency vector is perpendicular to the wavefronts of constant phase, coherent transfer function (the defocused pupil function) [Eq. (63)]. For
and is thus parallel to the optical direction cosine vector (cos 𝛼, cos 𝛽, axisymmetric optical system, the defocused pupil function is presented
cos 𝛾), so by Eq. (70)
√
𝑛 cos 𝛼 𝑛 cos 𝛽 𝑛 cos 𝛾 𝐻 (𝜌) = 𝑃 (𝜌)𝑒𝑗𝑘Δ𝑧 1−𝜆 𝜌{
2 2
𝑢𝑥 = , 𝑢𝑦 = , 𝑢z = (320)
𝜆 𝜆 𝜆 ( ) 1 𝜌 ≤ 𝑁𝐴 (327)
where n is the RI of the medium. Due to the fact that the frequency vec- 𝑃 (𝜌) = 𝑐 𝑖𝑟𝑐 𝑁𝐴𝜌 ∕𝜆 = 𝜆
0 else
tor is of length n/k, for any frequency vectors, the three frequency com- √
ponents (ux , uy , uz ) are not completely independent in the 3D Fourier where 𝜌 = 𝑢2𝑥 + 𝑢2𝑦 is the radial spatial frequency, P(𝜌) is the objective
space and are related by the following formula
√ pupil function with the cut-off frequency of 𝑁𝐴𝜆
. According to Hankel
𝑛 transform, the defocused PSF of coherent imaging system can be denoted
𝑢2𝑥 + 𝑢2𝑦 + 𝑢2𝑧 = (321)
𝜆 as
Eq. (321) suggests that the frequency vectors of plane waves in different √
ℎ(𝑟, 𝑧) = 𝑃 (𝜌)𝑒𝑗𝑘𝑧 1−𝜆2 𝜌2
𝐽0 (2𝜋𝑟𝜌)2𝜋𝜌𝑑𝜌 (328)
directions (spatial frequency) are the same, and they are all located on ∫𝜌
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 80. The illumination patterns, phase transfer functions, axial intensity derivatives and Fourier spectra, and final reconstructed phases under different illumination
patterns.

Note here we explicitly write the defocused PSF as a function of r and z Eq. (328) is quite similar to Eq. (324), which indicates that the 3D
in polar coordinates. This is because Eq. (328) actually represents the 3D coherent transfer function of the coherent optical field can only take
PSF of the imaging system, which describes the 3D complex amplitude non-zero value on the Ewald sphere in the 3D Fourier space (which is
distribution in the image space formed by an ideal point source. Taking understandable, because the 3D coherent transfer function itself is the
Fourier transform about z on both sides of Eq. (328), the 3D coherent Fourier transform of the 3D PSF,
( and the 3D PSF
) is actually a coherent
√( )
transfer function of the coherent optical field can be obtained 1
2
field defined in 3D space). 𝛿 𝜂 − 𝜆
− 𝜌2 represents the projec-

√ tion of the 2D pupil function P(𝜌) (coherent transfer function) onto the
√ ⎛ ( )2 ⎞ 3D Ewald sphere surface. The projection direction depends on the light
1
𝐻 (𝜌, 𝜂) = 𝑃 (𝜌)𝑒𝑗𝑘𝑧 1−𝜆2 𝜌2
𝑒−𝑗2𝜋𝑧𝑙 𝑑𝑧 = 𝑃 (𝜌)𝛿 ⎜𝑙 − − 𝜌2 ⎟ (329) propagation direction with reference to z-axis.
∫ ⎜ 𝜆 ⎟
⎝ ⎠
C. Zuo, J. Li and J. Sun et al.
Fig. 81. Description of a monochromatic plane wave in the 3D spatial and fre-
quency space.
Fig. 82. Generalized aperture P(u) as a 3D frequency spectrum limited by the
aperture angle on the Ewald sphere.
Note that in the above derivation, we assume that the light wave
propagates in vacuum. For the case that the light wave propagates in the
medium (e.g., when an oil immersion objective lens is used), the wave-
length in Eq. (328) should be replaced by the equivalent wavelength of
the light wave in the medium 𝜆/n, where n is the RI of the medium.
In addition, wave vector (𝑘𝑥 , 𝑘𝑦 , 𝑘𝑧 ) = 2𝜋(𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ) and the spatial fre-
quency vector (ux , uy , uz ) differ only by a factor of 2𝜋 and are therefore
used synonymously. The radius of Ewald sphere in the 3D wave vector
space is thus |𝐤| = 𝑛 2𝜆𝜋 (see Subsection 8.3 for details).
The aperture of the imaging system determines the maximum trans-
verse frequency of the spectrum, and thus, the lateral resolution limit.
For coherent imaging, the diffraction limit of the lateral resolution is
𝑁𝐴
𝜆
. However, when imaging thick objects, the limited aperture also
affects the diffraction limit of axial resolution (i.e., depth of field). As
shown in Fig. 82, the 2D complex pupil function is projected onto the
3D spherical surface, resulting in a segment of the Ewald sphere sur-
face. According to McCutchen [250], this segment of the Ewald sphere
is called the generalized aperture (i.e., 3D coherent transfer function), and
its inverse Fourier transform corresponds to the 3D PSF of the imaging
system. The generalized aperture is equivalent to a band-pass filter in 3D
Fourier space, whose lateral cut-off frequency corresponds to the √ coher-
𝑛− 𝑛2 −𝑁𝐴2
ent diffraction limit 𝑁𝐴
,
and the axial cut-off frequency is . ∫
𝜆 𝜆
The aperture angle 𝛼 denotes the largest cone of wave vectors that can
pass through the imaging lens. It should be noted that under the paraxial
approximation, the axial cut-off frequency of the generalized aperture
Optics and Lasers in Engineering 135 (2020) 106187
just corresponds to the normalized coefficient of the axial coordinate
used in Subsection 6.3.5.
8.3. Scattering potential representation and approximate conditions for 3D
objects
In the 2D theory of optical imaging, as described in Sections 4, a
thin object is usually represented as a 2D complex transmittance func-
tion composed of absorption component and phase component 𝑇 (𝑥, 𝑦) =
𝑎(𝑥, 𝑦) exp [𝑗𝜙(𝑥, 𝑦)]. The complex amplitude distribution of the transmit-
ted field is given by the product of the complex amplitude of the incident
illumination and the complex transmittance of the object. But when we
image a 3D object, how to model the object and describe its interaction
with the incident illumination? A straightforward idea is to extend the
2D definition, defining the thick object as a 3D complex transmittance
function T(x, y, z) so that the complex amplitude distribution of the
transmitted optical field can also be represented as the product of the
incident field and the object. This is the main content to be discussed in
this subsection.
Generally, the 3D phase imaging of thick object is also known as
ODT, which is first proposed by Wolf in 1969 [375]. Considering a 3D
sample with RI distribution n(r) illuminated by a monochromatic plane
wave travelling in 3D space (x, y, z), and the RI distribution of the sur-
rounding medium of the sample is nm . Note that n(r) is a complex distri-
bution, whose real and imaginary parts represent the RI and absorption
components of the sample, respectively. Since the RI of the object is dif-
ferent from that of the surrounding medium, the light propagation and
scattering obey the following inhomogeneous wave equation
[ 2 ]
∇ + 𝑘2 (𝐫 ) 𝑈 (𝐫) = 0 (330)
Note that the difference between Eq. (330) and Helmholtz equation
[Eq. (2)] in free space (in homogeneous medium) is that the wave num-
[ ]
ber 𝑘(𝐫 ) = 𝑘0 𝑛𝑚 +Δ𝑛(𝐫 ) is a variable related to the spatial distribution
of the RI. Here Δ𝑛(𝐫 ) = 𝑛(𝐫 ) − 𝑛𝑚 , k0 is the wave vector of light wave
in free space. U(r) represents the complex amplitude distribution of the
total 3D optical field in 3D space. Expanding the LHS of Eq. (330), we
obtain
( 2 )
∇ + 𝑘2𝑚 𝑈 (𝐫) = −𝑓 (𝐫 )𝑈 (𝐫 ) (331)
where 𝑘𝑚 = 𝑘0 𝑛𝑚 is the wave number in the medium. f(r) is defined as
the scattering potential function of the sample
[ ]
𝑓 (𝐫 ) = 𝑘20 𝑛(𝐫 )2 − 𝑛𝑚 2 (332)
Obviously, the scattering potential outside the sample is 0. In fact, the
scattering potential of the sample is equivalent to the complex RI dis-
tribution, so f(r) is often called the object function. The total field U(r)
can be written as the superposition of the incident (un-scattered) field
Uin (r), and the scattered field Us (r)
𝑈 (𝐫) = 𝑈𝑖𝑛 (𝐫) + 𝑈𝑠 (𝐫) (333)
while the incident field satisfies the Helmholtz equation in homogeneous
medium
( 2 )
∇ + 𝑘2𝑚 𝑈𝑖𝑛 (𝐫) = 0 (334)
Substituting Eq. (333) into Eq. (331) and using Eq. (334), the scattered
field satisfies the following equation
( 2 )
∇ + 𝑘2𝑚 𝑈𝑠 (𝐫) = −𝑓 (𝐫 )𝑈 (𝐫 ) (335)
The partial differential equation expressed in Eq. (335) cannot be di-
rectly solved, but based on Green’s method, its solution can be expressed
as the following integral form
( ) ( ) ( ) [ ( ) ( )]
𝑈𝑠 (𝐫 ) = 𝐺 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈 𝐫 ′ 𝑑 𝐫 ′ = 𝑓 𝐫 ′ 𝑈 𝐫 ′ ⊗ G(𝐫 ) (336)
where G(r) is the 3D Green’s function corresponding to Helmholtz equa-
tion in homogeneous medium, which is a outgoing spherical wave orig-
inates from the point of observation r
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

( 2 )
∇ + 𝑘2𝑚 G(𝐫 ) = −𝛿(𝐫 ) (337) Note that the complex phase of the scattered field 𝜙s (x) here is different
from the “phase” in the conventional sense. Instead, it represents the
( ) change of the complex phase of the incident field induced by the object.
exp 𝑗 𝑘𝑚 |𝐫 |
G(𝐫 ) = (338) Since 𝑈𝑠 (𝐫 ) = 𝑈 (𝐫 ) − 𝑈𝑖𝑛 (𝐫 ), substituting it into Eq. (343), the scattered
4𝜋|𝐫 |
optical field can be expressed
The convolution integral [Eq. (336)] suggests to decompose the source [ ] [ ]
term f(r)U(r), i.e., the product of the total field and the object function, 𝑈𝑠 (𝐫 ) = exp 𝜙𝑖𝑛 (𝐫 ) + 𝜙𝑠 (𝐫 ) − exp 𝜙𝑖𝑛 (𝐫 )
into the shifted and weighted sum of a delta function 𝛿(r). And the total [ ]{ [ ] }
= exp 𝜙𝑖𝑛 (𝐫 ) exp 𝜙𝑠 (𝐫 ) − 1
scattered field is the superposition of the scattered fields generated by { [ ] }
= 𝑈𝑖𝑛 (𝐫 ) exp 𝜙𝑠 (𝐫 ) − 1 (344)
these point sources 𝛿(r). We can further separate the source function
into two terms, which respectively represent the contributions of the Then, the complex phase of the scattered field can be represented as
incident field and the coupled correlation of the scattered field itself follows
( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) 𝑈 𝑠 (𝐫 ) 𝑈 (𝐫 )
𝑈𝑠 (𝐫) = 𝐺 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈𝑖𝑛 𝐫 ′ 𝑑𝐫 ′ + 𝐺 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈𝑠 𝐫 ′ 𝑑𝐫 ′ (339) 𝜙𝑠 (𝐫 ) = ln + 1 = ln (345)
∫ ∫ 𝑈𝑖𝑛 (𝐫 ) 𝑈𝑖𝑛 (𝐫 )
Though Eq. (339) provides a linearized relation between the scattered Based on above expressions, it can be proven that the solution to the
field and the scattering potential of the object, it is still quite challenging inverse scattering problem [Eq. (335)] can be written as the following
to invert it directly since Us (r) appears in the both sides of Eq. (339). integral form
However, if we assume that the light is scattered only once by the ob- [ ]
( ) ( ) ( ) | ( )|2
ject, i.e., the light is deflected only by a single interaction, a linearized 𝑈𝑖𝑛 (𝐫 )𝜙𝑠 (𝐫 ) = ∫ 𝐺 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈𝑖𝑛 𝐫 ′ + |∇𝜙𝑠 𝐫 ′ | 𝑑 𝐫 ′ (346)
| |
relation between the resultant the first-order scattered field, Us1 , and
the scattering potential of the object can be obtained Similarly, the first-order Rytov approximation is introduced to make this
problem linearized and solvable. When the phase gradient introduced
( ) ( ) ( )
𝑈𝑠1 (𝐫 ) = 𝐺 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈𝑖𝑛 𝐫 ′ 𝑑 𝐫 ′ (340) by the object is small so that the following slowly varying condition is
∫ satisfied
Us1 (r) represents the scattered field produced by the single scattering ( )2
2 𝜆
between the incident field and the object, which is only a portion of the 𝑛𝛿 ≫ ||∇𝜙𝑠 || (347)
2𝜋
total scattered field
where n𝛿 is the RI variation of the object within the range of one wave-
( ) ( ) ( )
𝑈𝐵 (𝐫) = 𝑈𝑠 (𝐫) − 𝐺 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈𝑠 𝐫 ′ 𝑑𝐫 ′ (341) length. Then, the contribution of the complex phase gradient term in
∫ Eq. (346) can be safely ignored for small n𝛿 , and Eq. (346) can be sim-
Eq. (340) suggests that the first-order scattered field of the entire object plified as
is the superposition of elemental scattered fields of all points that consti- ( ) ( ) ( )
tute the object, i.e, convolution of the Green’s function by the product of 𝑈𝑖𝑛 (𝐫 )𝜙𝑠 (𝐫 ) ≈ 𝑈𝐵 (𝐫) = 𝐺 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈𝑖𝑛 𝐫 ′ 𝑑𝐫 ′ (348)
∫
the object function and the incident field. As is clear here, the solution
It can be found that the complex phase of the scattered field can be inter-
requires the first-order scattered field Us1 (r) either to be a measurable
preted as the modulation to the incident field introduced by the object
quantity or can be obtained by other means. Two approximations are
𝜙s1 (x) ≈ Uin (r)/UB (r). The difference between Eq. (348) and Eq. (342) is
often used to determine Us1 (r):
only that the approximation imposed on the first-order scattered field
1) First-order Born approximation [374,375] Us1 (r). It is not difficult to prove that Rytov approximation can be re-
duced to Born approximation when the scattered field is weak or the
When the RI values of the sample and the medium are quite close,
total phase delay induced by the object is small [374]. Since the valid-
the light scattering should be very weak. In this case, it can be assumed
ity of the Rytov approximation is not dependent on the absolute phase
that the scattered field is negligible compared with the incident field
change introduced by the sample, but on the gradient of the RI within
Uin ≫ Us so that the second term in Eq. (341) can be ignored
the sample [374,384], this makes the Rytov approximation more suit-
( ) ( ) ( ) able to thick biological samples than the Born approximation [374,384–
𝑈𝑠 (𝐫) ≈ 𝑈𝑠1 (𝐫) = 𝐺 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈𝑖𝑛 𝐫 ′ 𝑑𝐫 ′ (342)
∫ 386].
Eq. (342) suggests that under the first-order Born approximation, the
total scattered field is approximated by the first-order Born scattered 8.4. Fourier diffraction theorem
field, and the high-order scattered field [the second term in Eq. (339)]
is ignored. In this case, there is a linear relationship between the object In the previous subsection, we learned that the 3D object is mod-
function and the total scattered field. It should be noted that the first- eled as the scattering potential satisfying the differential equation of
order Born approximation assumes that the objects are weak scattering. Eq. (335). Under Born or Rytov approximation, this equation can be
Numerical simulation verified that only if the absorption and total phase solved linearly, and the solution is represented as the integral form ex-
delay introduced by the object is small (typically less than 𝜋, ) the first- pressed by Eq. (340). In this subsection, we will discuss the physical
order Born approximation can produce accurate result. When the sample meaning behind Eq. (340) and how to use it to realize ODT for 3D ob-
is large or its RI is much higher than that of the surrounding medium, jects. First, rewriting the RHS of Eq. (340) into the convolution form
the first-order Born approximation will be no longer valid. In this case, [ ]
the first-order Rytov approximation is preferred. 𝑈𝑠1 (𝐫 ) = 𝐺(𝐫 ) ⊗ 𝑓 (𝐫 )𝑈𝑖𝑛 (𝐫 ) (349)

2) First-order Rytov approximation [374,382,383]
The first-order Rytov approximation assumes that the total field has a
complex phase function related to the scatted field, i.e., 𝑈 (𝐫 ) = exp [𝜙(𝐫 )]
[ ]
and 𝑈𝑖𝑛 (𝐫 ) = exp 𝜙𝑖𝑛 (𝐫 ) . The complex phase of the total field is the sum
of the complex phase of the incident field and the complex phase of the
scattered field
𝜙(𝐫 ) = 𝜙𝑖𝑛 (𝐫 ) + 𝜙𝑠 (𝐫 )
Taking Fourier transform on the both sides of Eq. (349) yields
{ }
𝑈̂ 𝑠1 (𝐤) = 𝐺̂ (𝐤) 𝑓̂(𝐤) ⊗ 𝑈̂ 𝑖𝑛 (𝐤) (350)
Here we use the wave vector 𝐤 = (𝑘𝑥 , 𝑘𝑦 , 𝑘𝑧 ) to represent the frequency
coordinates in 3D Fourier space, and its modulus is the wave number in
the medium |𝐤| = 𝑘𝑚 . In general, the incident field is a plane wave and
( )
can be expressed as 𝑈𝑖𝑛 (𝐫 ) = exp 𝑗 𝐤𝑖 ⋅ 𝐫 , where ki is the incident wave
(343) vector of plane wave. Since the image senor is a 2D detector array, we
C. Zuo, J. Li and J. Sun et al.
can only measure the field distribution on a plane. As we discussed at
the beginning of this section, the Fourier spectrum of the 3D coherent
optical field is highly redundant, and the complex amplitude distribu-
tion on a single plane is sufficient to fully characterize the whole 3D
field. It should be remembered that Green’s function in Eq. (350) serves
as a probing function which originates from the point r under consider-
ation, inside the volume. We first implement the 3D angular spectrum
decomposition (Fourier transform) of Green’s function [374], and then
take the 1D inverse Fourier transform with respect to z. Finally, assum-
ing that the detection plane is located at 𝑧 = 𝑧𝐷 , and only the forward
propagation component of the scattered field can be detected (neglect-
ing the back-scattering), we can get the following expression
′
𝑗𝑘𝑧 𝑧𝐷 ( )||
𝑒 tive 3D RI distribution. When the sample is illuminated by a plane wave
𝑈̂ 𝑠1 (𝐤⊥ , 𝑧𝐷 ) = −𝑗 𝑓̂ 𝐤⊥ − 𝐤𝑖⊥ , 𝑘𝑧 − 𝑘𝑖𝑧 ||
′
(351)
𝑘𝑧
′
| from a certain direction, the complex amplitude of the forward trans-
|𝑧𝐷 ≥0
Here the transverse frequency coordinate is defined as 𝐤⊥ = (𝑘𝑥 , 𝑘𝑦 ), and
′
√
we define a new quantity 𝑘𝑧 = 𝑘2𝑚 − 𝑘2𝑥 − 𝑘2𝑦 , which is a propagation
constant and only related to the transverse frequency kx , ky , to dis-
̂ total optical field in a certain plane U(r) is measured via interferometry
tinguish from the axial
( frequency) coordinate kz . 𝑈𝑠1 (𝐤⊥ , 𝑧𝐷 ) is the 2D
Fourier spectrum 𝑈̂ 𝑠1 𝑘 , 𝑘 , 𝑧𝐷 of the first-order scattered field at the
𝑥 𝑦
′
detection plane 𝑧 = 𝑧𝐷 . The exponential term 𝑒𝑗𝑘 𝑧𝐷 just corresponds to
the angular spectrum propagation kernel for a coherent complex field
[according to Eq. (326)], which accounts for the coordinate shift in the
z direction. Note that this exponential term will automatically vanish
if the measurement is performed at the nominal “in-focus” plane (z =
0). Because the Fourier spectrum of the 3D coherent field can only as-
sume non-zero values on the Ewald sphere and the object function on
the RHS is also defined in the 3D Fourier space, in order to show the
physical meaning behind Eq. (351) more clearly, we can also explicitly
write the LHS of Eq. (351) as a 3D function, which corresponds to the
projection of the 2D spectrum onto the 3D Ewald sphere
( )
exp 𝑗 𝑘𝑧 𝑧𝐷 ( )
𝑈̂ 𝑠1 (𝐤𝑠1 , 𝑧𝐷 ) = 𝑗 𝑓̂ 𝐤𝑠1 − 𝐤𝑖 (352)
𝑘𝑧
Eq. (352) is known as the Fourier diffraction theorem [374,375],
which relates the first-order scattered field at the detection plane
(𝑧 = (𝑧𝐷 ) to the scattering potential in the Fourier space. Here
√ )
𝐤𝑠1 = 𝑘𝑥 , 𝑘𝑦 , 𝑘z = 𝑘2𝑚 − 𝑘2𝑥 − 𝑘2𝑦 represents the 3D spatial frequency
coordinate of the first-order scattered field. It should be noted that the
the first-order scattered( field is
) defined on the Ewald sphere, and kz is
( )
actually dependent on 𝑘 , 𝑘 . 𝑓̂ 𝐤𝑠1 − 𝐤𝑖 on the RHS represents that
𝑥 𝑦
the spectrum of the scattered field ks1 is given by the object frequency
spectrum shifted by the frequency of the incident field ki
𝐠 = 𝐤𝑠1 − 𝐤𝑖 (353)
A consequence of Eq. (353) is the well-known Laue equation [220] with
ks1 being the scattered and ki being the incident wave vector. The Laue
equation results from the the assumption of single scattering. According
to the Laue equation the scattered wave vector ks1 can be determined
from the sum of an incident wave vector ki and a wave vector g from the
scattering object. In the far-field approximation, i.e., if evanescent waves
can be neglected, the transfer function of free space has also to be consid-
ered. This means that only the frequencies ks1 on the Ewald sphere can
propagate over distances large in comparison to the wavelength. Fig. 83
illustrates this in a graphical way. ks1 represents the Ewald sphere of
the first-order scattered field, ki is the incident vector. The propagating
contributions to the scattered field are limited to vectors on the Ewald
sphere. The wave vector g is the accessible frequency component of the
object, which is a shifted Ewald sphere passing through the origin with
center −𝐤𝑖 . Of course, here we ideally assume that we can capture the
transmitted and reflected first-order scattered field from both sides of
the object. Due to the technical layout, as for conventional microscopic
imaging in transmissive geometry, the reflected parts are mostly lost,
Optics and Lasers in Engineering 135 (2020) 106187
and only forward scattering can be received [see red circle in Fig. 83(a)].
In addition, we can change the direction of the wave vector ki by vary-
ing the illumination angle. If the object is illuminated from all possible
directions in 3D space (360∘ surround), it can be imagined that the max-
imum frequency coverage of the object becomes a sphere with a radius
of 2km . We call this large sphere Ewald limiting sphere in the object space
(which has a doubled radius compared with the Ewald sphere of the
coherent field), as illustrated in Fig. 83(c) [387].
8.5. 3D diffraction tomographic imaging under coherent illumination
The Fourier diffraction theorem [Eq. (352)] provides a powerful tool
for solving the inverse scattering problem and recovering the quantita-
mission component of the first-order scattered field is acquired. Taking
Fourier transform, we obtain a portion of object spectrum on the Ewald
sphere. The complex amplitude of the incident plane wave illumination
Uin (r) can be calibrated or predefined, and the complex amplitude of the
or phase retrieval approaches. Thus, the first-order scattered field can
be determined based on two approximations described in the previous
subsection
{
𝑈𝑠 (𝐫 ) = 𝑈 (𝐫 ) − 𝑈𝑖𝑛 (𝐫 ) ( ) 𝐵𝑜𝑟𝑛
𝑈 𝐵 (𝐫 ) ≈ (354)
𝑈𝑖𝑛 (𝐫 )𝜙𝑠 (𝐫 ) = 𝑈𝑖𝑛 (𝐫 ) ln 𝑈𝑈 ((𝐫𝐫)) 𝑅𝑦𝑡𝑜𝑣
𝑖𝑛
After obtaining Us1 (r), its planar 2D Fourier spectrum 𝑈̂ 𝑠1 (𝐤) is projected
onto a semi-spherical surface according to the Laue equation as depicted
in Fig. 83. By changing the incident angle of the illumination ki , the
resultant first-order scattered field can access different regions of the 3D
Fourier spectrum of the object, as illustrated in Fig. 85. With a series of
angle-dependent complex amplitude measurements, a certain portion of
the 3D Fourier spectrum of the object can be completed, and an estimate
of 𝑓̂(𝐤) can be obtained. Finally, an inverse 3D Fourier transform of the
resultant Fourier spectrum produces the 3D complex RI tomogram of
a sample. When the incident illumination covers all possible directions
( ± 90∘ ), the frequency coverage of the scattering potential is shown in
Fig. 85(a). It can be found that only a limited region of the spectrum
can be accessed, which results from the shifted Eward sphere through
the origin with center −𝐤𝑖 rotating ± 360∘ along the z axis. Although the
extension in the lateral direction reaches the border of Ewald limiting
sphere, a missing cone appears in the axial direction.
Additional limitations of the frequency coverage emerge from lim-
ited aperture angles or numerical apertures of the illumination optics
and the objective, respectively. For on-axis plane wave illumination, the
accessible object frequency just corresponds to the support of the gener-
alized aperture, with a lateral frequency extent Δ𝑢𝑥 = 𝑁𝐴 and an axial
√ 𝜆
𝑛− 𝑛2 −𝑁𝐴2
extent Δ𝑢𝑧 = 𝜆
, as shown in Fig. 82. The aperture angle corre-
sponding to the largest
( cone
) of wave vectors that can be illuminated and
𝑁𝐴
detected is 𝑎𝑟𝑐 sin 𝑛
. For matched illumination, i.e., the maximum
illumination angle allowed is limited to the maximum collection NA,
a torus-shaped structure in 3D frequency space is formed [Figs. 85(b)
and (d)]. Only those object frequencies can be transferred as a result
of the limited illumination and objective apertures. The resulting acces-
sible volume in frequency space represents the 3D transfer function of
tomographic imaging. It describes the possible object frequencies con-
tributing to the image formation, and can be used to assess both lateral
and axial resolution limit quantitatively. Incomplete illumination angle
leads to larger missing cone region, as shown in Fig. 85(c). The miss-
ing cone will reduce of the resolution of the reconstructed RI, especially
the axial resolution [376,378]. For partially coherent optical imaging,
the object is illuminated by a limited angular spectrum. Only the infor-
mation for a finite angular range is gathered. So, besides the low-pass
filtering by the imaging system, the finite illumination aperture has also
to be taken into account, which will be discussed later in Subsection 8.8.
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 83. Fourier-diffraction projection theorem. (a)-(b) Laue construction for determining a scattering wave vector under transmission and reflection; (c) the corre-
sponding Ewald limiting sphere under illuminations at all possible angles.

Fig. 85. The Fourier coverage of scattering potential with different transmissive
configurations. (a) Ideal free space with all possible incident angles; (b) finite
aperture imaging system with matched illuminations (𝑠 = 1); (c) finite aperture
imaging system with limited-angle illuminations (𝑠 = 0.7); (d) the corresponding
transverse/axial resolution limit in the case of finite aperture imaging system
with matched illuminations (𝑠 = 1).

Fig. 84. Illustration of the Fourier diffraction theorem under transmission and
reflection configurations. (a) Transmission configuration; (b) reflection config-
uration. axis of rotation (y). In order to alleviate the missing cone problem, itera-
tive non-negativity constraint can be applied to computationally fill the
missing information [384]. In addition, in the limit of small wavelengths
𝜆 → 0 (e.g., X-ray), the radius of the Ewald sphere km → ∞, the Fourier
Besides changing the illumination angle, we can also rotate the sam-
diffraction theorem converges to the Fourier slice theorem discussed in
ple to achieve larger spectrum coverage, which is similar to the case
Subsection 6.2.6 [374].
of conventional CT imaging. In this case, the illumination is fixed with
( )
normal incidence with respect to the detector 𝐤 = 0, 0, 𝑘𝑚 . Without
considering the aperture effect, the spectrum information once acquired
corresponds to a shifted hemisphere with vertex at the origin, as illus- 8.6. Different forms of 3D coherent transfer functions
trated in Fig. 86(a). The sample rotation is equivalent to spectrum ro-
tation. If the object is rotated about y-axis, the hemisphere can sweep In the 2D Fourier theory of optical imaging, as described in Section 4,
around the object spectrum along the sample-rotating axis (y), filling a the diffracted spectrum is given by the convolution of the incident spec-
horn torus-like shape in Fourier space, as shown in Fig. 86(b). Under trum with the object spectrum. As it turns out in the 3D treatment of
this setting, the maximum extension of the swept spectrum in kx and optical imaging, the reason for this is the linearity of diffraction in fre-
√
kz plane is 2𝑘𝑚 , as shown in Fig. 86(c). But notice that rotating the quency space, e.g., approximated by the first-order Born or Rytov ap-
sample only about one axis still results in missing-cone problem [376], proximation. When the effect of apecture is considered, and the inci-
as illustrated in Fig. 86(b). This results in directional blurring along the dent illumination is assumed to be perpendicular to the detection plane,
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 86. The Fourier coverage of scattering potential with sample rotation. (a)Information on the hemisphere surface can be obtained with a single illumination
direction; The surface of the sphere is oriented along z axis in the case of normal incidence; (b) rotating the object As a result of sample rotation, the hemisphere
sweep around the object spectrum along the sample-rotating axis (y), filling a horn torus-like shape inside the 3D spectrum of the object; (c) the 𝐤𝐱 − 𝐤𝐳 cross-section
of the region swept by the hemisphere.

Eq. (352) can be rewritten as

𝐻(𝑢𝑥 , 𝑢𝑦 )
𝑈̂ 𝑠1 (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ) = 𝑗 √( ) 𝑓̂(𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ) ⊗ 𝑈̂ 𝑖𝑛 (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 )
𝑛𝑚 2
2𝜋 𝜆
− 𝑢2𝑥 − 𝑢2𝑦
( √ )
𝐻(𝑢𝑥 , 𝑢𝑦 ) ( 𝑛 )2
̂ 𝑚
=𝑗 √( ) 𝑓 ( 𝑢 , 𝑢
𝑥 𝑦 𝑧, 𝑢 ) 𝛿 𝑢 𝑧 − − 𝑢 2 − 𝑢2
𝑥 𝑦
𝑛𝑚 2
𝜆
2𝜋 𝜆
− 𝑢 2 − 𝑢2
𝑥 𝑦

(355)

H(ux , uy ) is the 2D coherent transfer function with the limited aper-

ture function and propagation factor included [Eq. (63)]. It is plain to
see that the scattered spectrum is given by the convolution of the inci-
dent spectrum with the 3D frequency spectrum of the object, and there Fig. 87. The projection of 3D coherent transfer function onto the 2D plane.
is a linear relationship between the object function and the first-order
scattered field 𝑈̂ 𝑠1 (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ) = 𝐻3𝐷 (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ) ⋅ 𝑓̂(𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ), and they are
related by the following 3D transfer function
( √ )
𝐻(𝑢𝑥 , 𝑢𝑦 ) ( 𝑛 )2
𝑚
𝐻3𝐷 (𝑢𝑥 , 𝑢𝑦 , 𝑢𝑧 ) = √ ⋅ 𝛿 𝑢𝑧 − − 𝑢𝑥 − 𝑢𝑦 (356)
2 2
( )2 𝜆
𝑛𝑚
𝜆
− 𝑢 2 − 𝑢2
𝑥 𝑦

Note that we ignore the unimportant constant factor and imaginary unit
j in Eq. (356) [the extra j appears because the (real) amplitude and
(imaginary) phase information become to the (imaginary) absorption
and (real) RI in the scattering potential representation)]. Comparing
Eq. (356) with Eq. (329), it can be seen that the 3D coherent trans- Fig. 88. The comparison of Green’s functions used in angular spectrum propa-
fer function derived from the Fourier diffraction theorem is not co- gation and diffraction tomography. (a) Green’s functions corresponding to an-
incident with that derived directly from the angular gular spectrum propagation and diffraction tomography; (b) the Green’s func-
√spectrum
( )2
theory
tion of diffraction tomography has singular points in frequency domain (infinite
𝑛𝑚
in Subsection 8.2 due to the additional factor 𝑢𝑧 = 𝜆
− 𝑢2𝑥 − 𝑢2𝑦 . gain).
Therefore, according to the Fourier diffraction theorem, the 3D coher-
ent transfer function is not the direct and uniform projection of the 2D corresponds to the impulse response function [Eq. (13)] with the RS-I
coherent transfer function onto the 3D Ewald sphere, but is given by the type obliquity factor [𝐾 (𝜃) = cos 𝜃] [2,5]. Note that the propagation an-
weighted projection related to uz . Fig. 87 illustrates the transformation gle 𝜃 is a function of transverse coordinates in both original plane and
from 3D to 2D coherent transfer function representation. A homoge- diffraction plane. There is no field radiated along the transverse plane,
neous amplitude on the Ewald sphere, corresponding to, e.g., a spheri- as shown in Fig. 88(a). The 2D Fourier transform of the Green’s function
cal wave, leads to an amplitude distribution 1/cos 𝛼, increasing with the of the angular spectrum diffraction is
aperture angle 𝛼.
{ ( ) } { ( )}
The main reason for this discrepancy stems from the selection of exp 𝑗 𝑘𝑚 |𝐫 | 𝜕 exp 𝑗 𝑘𝑚 |𝐫 | ( )
Green’s function: When deriving the Fourier diffraction theorem, the ℱ𝑥,𝑦 cos 𝜃 = ℱ𝑥,𝑦 = exp 𝑗 𝑘𝑧 𝑧 (357)
𝑗 𝜆|𝐫 | 𝜕𝑧 𝑗 𝜆|𝐫 |
Green’s function used is a ideal diverging spherical wave with uniform
amplitude. When only forward scattering is considered, the Green’s The exponential is simply the phase shift of the corresponding plane
function is equivalent to the impulse response function [Eq. (13)] wave for propagation over a distance z. It can be found that the ad-
with the RS-II type obliquity factor [𝐾 (𝜃) = 1]. The Green’s function is ditional kz in the denominator is canceled due to the cosine obliquity
isotropic because it is just a half-space spherical wave. However, for the factor, which converges to the angle spectrum diffraction theory. In con-
angular spectrum diffraction formula [Eq. (30)], the Green’s function trast, for the diffraction tomography representation, there is an extra kz
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Table 7 8.7. 3D partially coherent transfer functions, 6D TCC and 3D WOTF

Apodization functions of different design conditions
(𝜌 = cos 𝜃∕𝜆) In the previous subsections, we have discussed the following linear
Design conditions Apodization function relationship between the object function and the first-order scattered
√ field in 3D coherent imaging
Sine condition 𝑃 (𝜃) = 𝑃 (𝜌) cos 𝜃
Herschel condition 𝑃 (𝜃) = 𝑃 (𝜌) ( ) ( ) ( )
√
Lagrangian condition 𝑃 (𝜃) = 𝑃 (𝜌) 𝜃∕sin 𝜃 𝑈𝑠1 (𝐫 ) = ℎ 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈𝑖𝑛 𝐫 ′ 𝑑 𝐫 ′ (359)
( √ )3 ∫
Helmholtz Condition 𝑃 (𝜃) = 𝑃 (𝜌) 1∕ cos 𝜃
where h(r) is the 3D coherent PSF, whose Fourier transform is the gen-
eralized aperture 𝐻 (𝛒, 𝑙) = ℱ {ℎ(𝐫 )}. In order to simplify the derivation,
in the denominator. The magnitude of the Fourier transform shows a we introduce the first-order Born approximation Us1 (r) ≈ Us (r), the total
singularity at 𝑘𝑧 = 0, i.e., at a spatial frequency of 𝑘⊥ = 𝑘𝑚 [Fig. 88(b)]. field can be expressed as
It stems from the plane wave component propagating perpendicular to
z-axis (cf. the angular spectrum of a spherical wave). 𝑈 (𝐫) = 𝑈𝑖𝑛 (𝐫 ) + 𝑈𝑠 (𝐫)
Though both widely used among the literatures, the 3D coherent ( ) ( ) ( )
transfer functions obtained from two different derivations are inconsis- = 𝑈𝑖𝑛 (𝐫 ) + ℎ 𝐫 − 𝐫 ′ 𝑓 𝐫 ′ 𝑈𝑖𝑛 𝐫 ′ 𝑑 𝐫 ′
∫
tent with each other. Under the paraxial approximation cos 𝜃 ≈ 1, the ( ) ( ) ( )
weighting factor in Eq. (358) vanishes, and then the two forms are uni- = ℎ 𝐫 − 𝐫 ′ 𝑇 𝐫 ′ 𝑈𝑖𝑛 𝐫 ′ 𝑑 𝐫 ′ (360)
∫
fied. But in the cases of high-resolution 3D imaging with high-NA illu-
mination and objective, the effect of uz cannot be ignored. How to dis- [ ]
where 𝑇 (𝐫 ) = 1 + 𝑓 (𝐫 ) = 1 + 𝑘20 𝑛(𝐫 )2 − 𝑛𝑚 2 is defined as the 3D complex
tinguish the correct one for a given situation is still a controversial issue
transmittance function of the object. Based on this complex transmit-
[388,389]. In the literature on the theory of image formation in the con-
tance expression, the 3D coherent imaging and 2D coherent imaging
focal fluorescence microscope, the 3D coherent transfer function with a
can be formally unified. In the following, we follow the path mirroring
large NA is often evaluated based on Debye integral [250,252,390,391].
that in Subsection 6.3 to derive the 3D transfer functions under partially
In the diffraction computation according to the Debye approximation,
coherent illuminations. When an 3D object is illuminated by a partially
each point on the aperture plane is taken as the source of a plane wave
coherent field with the CSD defined as WS (r1 , r2 ), the CSD of the trans-
with amplitude, phase and polarization prescribed by this point. The
mitted field can be expressed as
final image field is the superposition of plane wavelets within the geo-
metric cone (solid angle) limited by the aperture. For the low-NA case, ( ) ( ) ( ) ( )
𝑊𝑂 𝐫1 , 𝐫2 = 𝑊𝑆 𝐫1 , 𝐫2 𝑇 𝐫1 𝑇 * 𝐫2 (361)
the pupil function of objective is just the generalized aperture function
[Eq. (329)]. For larger NAs (NA > 0.75), the effect of the apodiza- According to Van Cittert-Zernike theorem, the CSD of the illumination
tion function become pronounced and must be included in the imaging before reaching the object can be written as
model. The apodization function P(𝜃) is not equal to the pupil function
P(𝜌) for a high NA objective: pupil function P(𝜌) represents the ray den- ( ) ( )
𝑊𝑆 𝐫1 , 𝐫2 = 𝑊𝑆 𝐫1 − 𝐫2 = 𝑆 (𝐮)𝑒𝑗2𝜋𝐮⋅(𝐫1 −𝐫2 ) 𝑑𝐮 (362)
sity in the transverse plane while the apodization function P(𝜃) gives ∬
the ray density over the converging wavefront [76,359]. Apodization
functions depend on the transmission coefficients at various interfaces Note that here u represents the 3D frequency vector (𝝆, 𝜂). Similar to
of an imaging system and on spatial filters inserted in the path of an the generalized aperture, S(u) refers to the incoherent generalized source
imaging system. The relationship of an apodization function to the cor- defined on the Eward sphere (in 6f imaging system the source is located
responding pupil function can be represented as a weighted projection at the conjugated plane of the objective pupil)
g(𝜃): √
√ ⎛ ( )2 ⎞
1
⎛ ( )2 ⎞ 𝑆 (𝛒, 𝜂) = 𝑆 (𝛒)𝛿 ⎜𝜂 − − 𝛒2 ⎟ (363)
𝐻 (𝜌, 𝜂) = 𝑃 (𝜌)𝑔 (𝜃)𝛿 ⎜𝜂 −
1
− 𝜌2 ⎟ (358) ⎜ 𝜆 ⎟
⎜ 𝜆 ⎟ ⎝ ⎠
⎝ ⎠
The commonly used apodization functions are provided in Table 7. For a After 6D Fourier transform, the transmitted light field needs to multiply
commercial objective, the sine condition is usually obeyed in design pro- the 3D coherent transfer function H(u) [Eq. (358)] twice at the pupil
cesses so that a perfect image of a thin object can be obtained within the plane, and then the CSD in the image space can be obtained after inverse
field of view of an objective. Such an imaging system exhibits 2D trans- Fourier transform
verse space invariance, which is called the aplanatic imaging system. ( ) ( ) ( ) ( )
𝑃 (𝜃) = 𝑃 (𝜌) cos 𝜃 indicates that the ray density over the angle of con- 𝑊𝐼 𝐫1 , 𝐫2 = ∫ ∫ 𝑊̂ 𝑂 𝐮1 , 𝐮2 𝐻 𝐮1 𝐻 * 𝐮2 𝑒𝑗2𝜋 (𝐮1 𝐫1 +𝐮2 𝐫2 ) 𝑑 𝐮1 𝑑 𝐮2 (364)
vergence is decreased with the aperture angle. The Herschel condition
𝑃 (𝜃) = 1 is also called the uniform angular condition. The apodization in which H(u1 )H∗ (u2 ) is the 3D mutual coherent transfer function.
function is completely consistent with the generalized aperture function Equivalently, it can also be written in the convolution form in the spatial
[Eq. (329)]. In other words, the ray density is constant over the range of domain
the angle of convergence. Under the Helmholtz condition, the apodiza-
( ) ( ) ( ) ( )
tion function increases with the aperture angle, which is similar to the 𝑊𝐼 𝐫1 , 𝐫2 = 𝑊𝑂 𝐫 ′ 1 , 𝐫 ′ 2 ℎ 𝐫1 − 𝐫 ′ 1 ℎ∗ 𝐫2 − 𝐫 ′ 2 𝑑 𝐫 ′ 1 𝑑 𝐫 ′ 2 (365)
∬
trend of the 3D coherent transfer function of Fourier diffraction theorem
[Eq. (356)]. However, the Helmholtz condition is generally only avail- where h(r1 )h∗ (r2 ) is the 3D mutual PSF. The image intensity that we can
able for reflective lenses, and the design of most microscopic objective captured is given by the values “on the diagonal” of the CSD function
follows the sine condition. In the rest part of this tutorial, the 3D coher-
( ) ( ) ( )
ent transfer function in Eq. (329) (Herschel condition) is used for sim- 𝐼 (𝐫 ) = 𝑊𝐼 (𝐫, 𝐫 ) = ∬ 𝑊𝑆 𝐫1 , 𝐫2 𝑇 𝐫1 𝑇 𝑠∗ 𝐫2
( ) ∗( ) (366)
plicity in the derivation of the 3D partially coherent transfer functions. ℎ 𝐫 − 𝐫1 ℎ 𝐫 − 𝐫2 𝑑 𝐫1 𝑑 𝐫2
But it should be noted that in high-NA microscopy applications, the sine
condition should be considered to produce more physically meaningful This formula can be further simplified by substituting Eq. (362) into
results. Eq. (366)
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

| ( ) ( ) |2 𝐻𝐴 (𝐮) = 𝑊 𝑂𝑇 𝐹 (𝐮) + 𝑊 𝑂𝑇 𝐹 ∗ (−𝐮) (376)

𝑆 (𝐮)|| 𝑇 𝐫 ′ ℎ 𝐫 − 𝐫 ′ 𝑒𝑗 2𝜋𝐮𝐫 𝑑 𝐫 ′ || 𝑑𝐮
′
𝐼 (𝐫 ) =
∫ | ∫ |
𝐻𝑃 (𝐮) = 𝑊 𝑂𝑇 𝐹 ∗ (𝐮) − 𝑊 𝑂𝑇 𝐹 (−𝐮) (377)
≡ 𝑆 (𝐮)𝐼𝐮 (𝐫 )𝑑𝐮 (367)
∫ Note that here 3D ATF and 3D PTF is then given by the real and imagi-
nary components of 3D WOTF respectively, which is now in coincidence
Eq. (367) suggests that the final intensity that can be captured at the im-
with the 2D case. The reason for this lies in the additional imaginary
age plane can be regarded as the intensity (incoherence) superposition
unit in the projection of the 2D aperture onto the 3D Eward sphere, as
of coherent partial images Iu (r) arising from all points over the incoher-
described in Eq. (355).
ent generalized source function. This is similar to the 2D case discussed
in Subsection 6.3.1. Substituting Eq. (362) into Eq. (367) and rewriting
8.8. 3D PTFs for coherent, incoherent, and partially coherent imaging
it in the form of Fourier integral gives
Next, we analyze the 3D transfer functions of an axisymmetric opti-
𝐼 (𝐫 ) = 𝑆(𝐮)𝑇̂ (𝐮1 )𝑇̂ ∗ (𝐮2 )𝐻(𝐮 + 𝐮1 )𝐻 ∗ (𝐮 + 𝐮2 )𝑒𝑗 2𝜋𝐫 (𝐮1 −𝐮2 ) 𝑑 𝐮1 𝑑 𝐮2 𝑑𝐮 cal system under different illumination conditions. First, under coherent
∭
(368) illumination (coherence parameter s → 0), the light source degenerates
into an ideal on-axis point radiator
Separating the contribution of the specimen and imaging system leads ( )
1
to the notion of the 3D TCC 𝑆 (𝛒, 𝜂) = 𝛿(𝛒)𝛿 𝜂 − (378)
𝜆
( ) √
𝑇 𝐶 𝐶 𝐮1 , 𝐮2 = 𝑆(𝐮)𝐻(𝐮 + 𝐮1 )𝐻 ∗ (𝐮 + 𝐮2 )𝑑𝐮 (369) where 𝜌 = 𝑢2𝑥 + 𝑢2𝑦 represents the radial spatial frequency coordi-
∬
nate and 𝜂 represents the axial spatial frequency coordinate. Taking
Noted that the 3D TCC is actually a 6D function. Based on the 3D TCC, Eq. (378) into the 3D WOTF
Eq. (368) can be written as ( )
1
( ) 𝑊 𝑂𝑇 𝐹 (𝜌, 𝜂) = 𝐻 𝜌, 𝜂 + (379)
𝐼 (𝐫 ) = 𝑇̂ (𝐮1 )𝑇̂ ∗ (𝐮2 )𝑇 𝐶 𝐶 𝐮1 , 𝐮2 𝑒𝑗 2𝜋𝐫 (𝐮1 −𝐮2 ) 𝑑 𝐮1 𝑑 𝐮2 (370) 𝜆
∬ It is shown that under coherent illumination, the 3D WOTF just corre-
The above expressions are similar to those derived from 2D partial co- sponds to the shifted generalized aperture, i.e., the shifted 3D coherent
herence imaging, except that the dimension of the TCC increases from transfer function H(u) [Eq. (358)] passing through the origin. This is
four to six. It can be calculated from the overlapping region of the gener- not difficult to understand because the linearization of 3D imaging is
alized source and two shifted generalized aperture (three hemispherical derived based on the first-order Born approximation (e.g., the Fourier
shells). However, the calculation requires prohibitively large computa- diffraction theorem), indicating that the 3D coherent transfer function
tion time for the 6D integral of the TCC. Considering the real part and implies the “weak object” approximation. Ideally, the generalized aper-
the imaginary part of the 3D complex transmission function separately, ture is a real function, the even part and the odd part of the 3D WOTF
i.e., the phase component P(r) and the absorption component A(r) (note correspond to the ATF and PTF, respectively.
that the real part of the 3D object function corresponds to the phase ( ) ( )
1 1
𝐻𝐴 (𝜌, 𝜂) = 𝐻 𝜌, 𝜂 + + 𝐻 𝜌, −𝜂 − (380)
component, instead of absorption as in the 2D case) 𝜆 𝜆
[ ] ( ) ( )
𝑇 (𝐫 ) = 1 + 𝑓 (𝐫 ) = 1 + 𝑘20 𝑛(𝐫 )2 − 𝑛2𝑚 𝐻𝑃 (𝜌, 𝜂) = 𝐻 𝜌, 𝜂 +
1
− 𝐻 𝜌, −𝜂 −
1
(381)
𝜆 𝜆
= 1 + 𝑃 (𝐫 ) + 𝑗 𝐴 (𝐫 ) (371)
They are two back-to-back Eward hemispherical shells intersecting at
Under the first-order Born approximation, the self-interference of the the origin. The 3D ATF is even symmetric about the 𝜂-axis with the
scattered light can be neglected, and the mutual spectrum of the object maximum value at the origin; while the 3D phase transfer function is
can be approximated as odd symmetric about the 𝜂-axis with zero at the origin, indicating that
( ) ( ) ( )[ ( ) ( )] a homogeneous object produces no phase contrast.
𝑇̂ (𝐮1 )𝑇̂ ∗ (𝐮2 ) = 𝛿 𝐮1 𝛿 𝐮2 + 𝛿 𝐮2 𝑃̂ 𝐮1 + 𝑗 𝐴̂ 𝐮1 + In the case of incoherent illumination, the generalized source is equal
( )[ ( ) ( )]
𝛿 𝐮1 𝑃̂ 𝐮2 − 𝑗 𝐴̂ 𝐮2 (372) to or even larger than the generalized aperture 𝑠 = 𝑁 𝐴𝑖𝑙𝑙 ∕𝑁 𝐴𝑜𝑏𝑗 ≥ 1,
and the 3D WOTF becomes the autocorrelation of 3D coherent transfer
Substituting Eq. (372) into Eq. (371) and using the Hermitian symmetry function
property of TCC yields the intensity of the partially coherent image for ( ) ( )
𝑊 𝑂𝑇 𝐹incoh (𝐮) = 𝐻 𝐮′ + 𝐮 𝐻 * 𝐮′ 𝑑 𝐮′ (382)
a weak object ∬
{ }
[ ] Note that it is always a real function (even under defocusing), which
𝐼 (𝐫 ) = 𝑇 𝐶 𝐶 (𝟎, 𝟎) + 2 Re 𝑇 𝐶 𝐶 (𝐮, 𝟎) 𝑃̂ (𝐮) + 𝑗 𝐴̂ (𝐮) 𝑒𝑗 2𝜋𝐫 𝐮 𝑑𝐮 (373)
∫ means the PTF is always 0. When the apodization function in the gen-
eralized aperture/3D coherent transfer function H(u) [Eq. (358)] is ig-
where TCC(u, 0) is the linear part of the TCC, which is the 3D WOTF
nored (satisfying Herschel condition), the transfer function can be ex-
( ) ( ) ( ) pressed as [391]
𝑊 𝑂𝑇 𝐹 (𝐮) ≡ 𝑇 𝐶 𝐶 (𝐮, 𝟎) = 𝑆 𝐮′ 𝐻 𝐮′ + 𝐮 𝐻 * 𝐮′ 𝑑 𝐮′ (374) [ ( )]
∬ 4 1 2 cos 𝛼
𝑊 𝑂𝑇 𝐹incoh (𝜌,
̄ 𝜂)
̄ = arccos +1 (383)
Taking Fourier transform on Eq. (373), and separating the contributions 𝜋𝐾 𝑀 |𝜂|
̄
√ √
of phase component P(r) and absorption component A(r) gives
where 𝐾 = 𝜌̄2 + 𝜂̄ 2 , 𝑀 = 𝐾2||𝜌|𝜂|
̄
1 − 𝐾4 , and 𝛼 is the aperture angle of
2
̄
𝐼̂(𝐮) = 𝐼0 𝛿(𝐮) + 𝐻𝐴 (𝐮)𝐴̂ (𝐮) + 𝐻𝑃 (𝐮)𝑃̂ (𝐮) (375) the imaging system. Note here the lateral and axial frequencies 𝜌, ̄ 𝜂̄
are both normalized by 1/𝜆. Fig. 89 shows the formation of the 3D
where I0 is the background intensity. It now becomes obvious that the
incoherent OTF. Because of the nature of the autocorrelation func-
image contrast due to the absorption and phase are decoupled and lin-
tion, the support of the 3D incoherent OTF is a torus-shaped region
earized. Because the intensity is always real, the real and even compo-
of Fourier space. The maximum accessible √ object frequencies in free
nent of WOTF(u) corresponds to the partially coherent 3D ATF HA (u),
1− 1−𝑁𝐴2
and the imaginary and odd component of WOTF(u) corresponds to the space correspond to Δ𝑢𝑥 = 𝑁𝐴
𝜆
and Δ𝑢𝑧 = 𝜆
, which gives a re-
3D PTF HP (u) of the object, respectively. minder of the normalized coordinate system used in Subsection 4.2 and
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Expanding the exponential term in Eq. (389) based on Euler formula,

we can get the following relationship between 3D transfer functions and
defocused 2D transfer functions [217]

𝐻𝐴 (𝜌, 𝑧) = 𝐻𝐴 (𝜌, 𝜂) cos (2𝜋𝑧𝜂)𝑑𝜂 (390)

∫

𝐻𝑃 (𝜌, 𝑧) = 𝐻𝑃 (𝜌, 𝜂) sin (2𝜋𝑧𝜂)𝑑𝜂 (391)

∫
It can be verified that, without considering the apodization function in
the generalized aperture/3D coherent transfer function (satisfying the
Herschel condition), the 3D ATF and PTF derived in this subsection
can be directly converted to the 2D defocused ATF and PTF derived
in Subsection 4.3 [Eqs. (92) and (93)].

8.9. Diffraction tomographic (3D phase) imaging under partially coherent

illuminations

As introduced at the beginning of this section, the widely applied

ODT techniques [375,376,378,379] usually require specially designed
holographic microscopes working with a coherent light source that se-
Fig. 89. Transverse and axial cross-sections of the 3D incoherent WOTF, repre- quentially illuminates the object from different sides by using a scanning
sented in normalized coordinates. module [384,392–394] or rotating the object mechanically [395,396].
However, such coherent ODT techniques recover the 3D RI by assem-
Subsection 6.3.4. It can be found the normalization coefficients are as- bling numerous sections of the object spectrum obtained for each il-
sociated with the maximum extent of the 3D incoherent OTF. And the lumination directions. Moreover, coherent ODT techniques suffer from
lateral and axial extents become 0.5 and 2 in the normalized coordinate speckle noise and parasitic interference due to the coherent illumina-
system (𝜌,
̄ 𝜂),
̄ as clearly illustrated in Fig. 89. tion. In this subsection, we will discuss another way to achieve 3D
Under partially coherent illuminations, the transfer functions of an diffraction tomography by extending the idea of TIE, namely TIDT. Sim-
axial-symmetric aberration-free imaging system are given by ilar to QPI techniques based on TIE, TIDT does not require coherent
illumination and interferometry, it only captures the through-focus in-
𝐻𝐴 (𝜌,
̄ 𝜂)
̄ = 𝑊 𝑂𝑇 𝐹 (𝜌,
̄ 𝜂)
̄ + 𝑊 𝑂𝑇 𝐹 (𝜌,
̄ −𝜂)
̄ (384) tensity images and directly recover the 3D RI distribution of the object
by 3D deconvolution based on the 3D PTF. The basic idea of TIDT is sim-
𝐻𝑃 (𝜌,
̄ 𝜂)
̄ = 𝑊 𝑂𝑇 𝐹 (𝜌,
̄ 𝜂)
̄ − 𝑊 𝑂𝑇 𝐹 (𝜌,
̄ −𝜂)
̄ (385) ilar to the widefield 3D fluorescence deconvolution microscopy, except
that the 3D transfer function used here is the 3D partially coherent PTF,
where 𝑊 𝑂𝑇 𝐹 (𝜌,
̄ 𝜂)
̄ is represented as
instead of incoherent 3D OTF [178]. It can be viewed as an extension of
√ (√ ) ⎛ ⎞ TIE phase retrieval technique, which generalizes the idea of 2D “trans-
𝜆 𝜌̄2 𝜎 𝐾2 𝐾 2 𝜎2 𝐾2 𝜂̄ 2 ⎜ 𝐾𝜎 ⎟
𝑊 𝑂𝑇 𝐹 (𝜌,
̄ 𝜂)
̄ = 1− − + 1− − arccos ⎜ √ / ⎟ port of intensity at a 2D plane” into “transport of intensity through a
2𝜋 𝐾 2 𝜂̄ 4 𝜂̄ 2 4 2𝐾 ⎜ 𝜂̄ 1 − 𝐾 2 4 ⎟
⎝ ⎠ 3D volume” to realize diffraction tomography. Compared with the ob-
(386) ject rotation or beam scanning, the data acquisition based on the axial
scanning not only greatly simplifies the system configuration, but also
when 0 < 𝜌̄ < 𝜌̄𝑃 − 𝜌̄𝑆 :
( ) √ √ √
is completely compatible with conventional wide-field transmission mi-
√ ( )2 √ ( )2
𝜂̄ 𝜂̄ croscope platforms.
𝜎= − 1 − 𝜌̄2𝑆 , 𝑖𝑓 1 − 𝜌̄2𝑆 − 1 − 𝜌̄𝑃 − 𝜌̄𝑆 ⩽ 𝜂̄ ⩽ 1 − 𝜌̄2𝑆 − 1 − 𝜌̄𝑃 + 𝜌̄𝑆
𝜌̄ 2
Fig. 90 illustrate the difference between the traditional coherent ODT
(387) and the partially coherent TIDT techniques. Traditional coherent ODT
when 0 < 𝜌̄ < 𝜌̄𝑃 − 𝜌̄𝑆 : techniques recover the 3D RI by assembling numerous of sections of the
( √ ) √ √ √ √ object spectrum obtained for each incident angle. With a series of angle-
𝜂̄ 𝜂̄ ( )2
𝜎= − 1 − 𝜌̄2𝑆 , 𝑖𝑓 1 − 𝜌̄2𝑆 − 1 − 𝜌̄𝑆 − 𝜌̄ ⩽ 𝜂̄ ⩽ 1 − 𝜌̄2𝑆 − 1 − 𝜌̄2𝑃 dependent complex amplitude measurements, a certain portion of the
𝜌̄ 2
( √ ) √ √ √ object spectrum can be filled, which produces an estimate of 3D com-
𝜂̄ 𝜂̄ ( )2 √
𝜎= − − 1 − 𝜌̄2𝑃 , 𝑖𝑓 1 − 𝜌̄2𝑆 − 1 − 𝜌̄2𝑃 ⩽ 𝜂̄ ⩽ 1 − 𝜌̄𝑃 − 𝜌̄ − 1 − 𝜌̄2𝑆 plex RI of the sample [Fig. 90(a)]. In contrast to the coherent ODT, TIDT
𝜌̄ 2
(388)
where 𝜌̄𝑆 and 𝜌̄𝑃 are both normalized by 1/𝜆 so that 𝜌̄𝑆 = 𝑁 𝐴𝑖𝑙𝑙 , 𝜌̄𝑃 =
𝑁 𝐴𝑜𝑏𝑗 . In fact, the partially coherent 3D WOTF was first derived by
Streibl [178] under the paraxial approximation as early as 1985, and
then it was further summarized and applied to 3D microscopic imaging
by Sheppard [179] based on the TCC representation. The expression of
the transfer function under the paraxial approximation is much simpler.
Until recently, the analytical expression of the 3D WOTF under the non-
paraxial condition was derived, and has been successfully applied to the
high-resolution TIDT [214,381,388].
Finally, it should be noted that we can directly obtain the 2D de-
focused transfer functions from the 3D transfer functions. The Fourier
transform of 3D intensity and the 2D intensity at different depths can
be related by taking the Fourier transform about z [177] Fig. 90. Comparison of two different diffraction tomography strategies under
coherent illumination (the conventional ODT) and partially coherent illumina-
𝐼 𝑧 (𝐫 ) = 𝐼 (𝐮)𝑒−𝑗2𝜋𝑧𝜂 𝑑𝜂 (389) tion (TIDT). (a) Coherent illumination; (b) partially coherent illumination.
∫
C. Zuo, J. Li and J. Sun et al.
can be regarded as its “parallelized” version due to the use of partially
coherent illumination. The sample is simultaneously illuminated by a
quasi-monochromatic and spatially incoherent extended source from all
possible directions allowed by the condenser aperture of the wide-field
microscope [Fig. 90(b)]. The raw data is the through-focus intensity
stack obtained by optical refocusing (axial scanning). Thus, TIDT is in-
herently faster and more efficient in terms of spectrum coverage and, in
practice, simplifies the measurement and the reconstruction process. An
additional advantage is that the low spatial coherent illumination avoids
speckle noise, which plagues coherent ODT and other coherent imaging
methods. It should be noted that though the fundamentals of TIDT were
established by Streibl [178] more than thirty years ago, only recently its
experimental implementation has been demonstrated [214,380,381].
In Subsection 8.7 and Subsection 8.8, we have learned that for
weakly scattering object under partially coherent illumination, the 3D
Fourier transform of the stack of intensity distributions at the image
space is given by
𝐼̂(𝐮) = 𝐼0 𝛿(𝐮) + 𝐻𝐴 (𝐮)𝐴̂ (𝐮) + 𝐻𝑃 (𝐮)𝑃̂ (𝐮) (392)
It should be noted that in Eq. (392), the phase (RI) information and the
intensity (absorption) information of the object are coupled together,
i.e., both phase contrast and amplitude contrast contribute to the mea-
sured intensity. In order the recover the quantitative RI information for
objects with nonnegligible absorption, we need to decouple it from the
contribution arising from the intensity information. There are four fea-
sible solutions:
1) Recording an additional set of intensity stacks with opposing il-
lumination or via object rotation over 180∘ : the symmetries of 3D
ATF and PTF allow the phase component to be separated by subtrac-
tion of the two intensity stacks recorded with opposing illumination
or via object rotation over 180∘ because the absorption contrast is
an even function about each scatterer. This is similar to TIE phase
retrieval based on central finite difference with the defocus distances
of the two images equal and opposite relative to the central in-focus
image.
2) Collecting two or more datasets of the same object with differ-
ent pupil functions: since the corresponding 3D transfer functions
are different, it should be possible to separate the phase contrast
and amplitude contrast by solving the corresponding linear equa-
tions [178].
3) Assuming the phase-attenuation duality analogous to 2D phase
imaging: the absorption is assumed to be proportional to RI 𝑃̂ (𝐮) ≈
𝜀𝐴̂ (𝐮), then Eq. (392) can be simplified as [380,381]
[ ]
𝐼̂(𝐮) = 𝐵 𝛿(𝐮) + 𝑃̂ (𝐮) 𝐻𝑃 (𝐮) + 𝜀𝐻𝐴 (𝐮) (393)
where 𝜀 is a constant parameter, which is usually empirically ad-
justed to optimize the reconstruction quality [381]. When the sample is
a pure phase object, 𝜀 = 0, and the 3D RI reconstruction boils down to
the following one-step 3D deconvolution
𝐼̂(𝐮)𝐻𝑃∗ (𝐮)
𝑃̂ (𝐮) = (394)
|𝐻 (𝐮)|2 + 𝛽
| 𝑃 |
where 𝛽 is the Tihkonv regularization parameter determined by the SNR
of the raw data. The support and response of the 3D PTF are directly
associated with the resolution and SNR of the 3D RI reconstruction.
According to Eq. (385), we can analyze the 3D PTF distribution of a
perfectly aligned microscope with rotationally symmetric illumination.
Figs. 91(a)-(d) display the 𝜌̄ − 𝜂̄ sections of the 3D PTF corresponding to
an objective with 𝑁𝐴 = 1.4 and circular illumination aperture with dif-
ferent coherence parameters 𝑠 = 0.3, 0.6 and 0.9. As can be seen, smaller
coherence parameters provide better low-frequency response at an ex-
pense of smaller spectral coverage, leading to compromised lateral and
axial resolution. As the coherence parameter increases, the spectral cov-
erage is extended, and the response at both low and high frequencies
Optics and Lasers in Engineering 135 (2020) 106187
also gradually increases. If the coherence parameter further increases to
∼ 1, the spectral coverage reaches the maximum (incoherent diffraction
limit). However, phase contrast progressively vanishes, suggesting that
the phase information can hardly be transferred into intensity when the
illumination NA is large. In other words, the intensity image gives no
phase contrast under incoherent illumination. Thus, there is an inherent
tradeoff between phase contrast and imaging resolution in TIDT with
conventional circular illumination apertures. Note that these observa-
tions are analogous to the case of the 2D QPI based on the TIE phase
retrieval. Smaller coherence coefficients correspond to stronger phase
contrast especially for low spatial frequencies, but the resolution is lim-
ited. Larger coherence coefficients provide higher theoretical resolution
limit, while the poor phase contrast compromises the SNR for phase re-
construction. Therefore, considering both of these two aspects, s should
be set between 0.4 and 0.7 to guarantee better 3D phase imaging per-
formance. It should be also noticed that, regardless of the illumination
aperture, there always exists a missing cone frequency region around the
𝜂̄ axis, which is a well-known common problem of the ODT techniques
for wide-field transmission configuration.
There are two possible strategies to overcome the above issue and
improve the performance of TIDT. One is to optimize the source distribu-
tion and improve the homogeneity of the PTF [397]. A similar approach
has been used to optimize the illumination pattern for the TIE phase
retrieval [211], as discussed in Subsection 7.6. It has been found that
Gaussian illumination is better than traditional uniform circular illumi-
nation for its higher homogeneity and less low-value regions in the 3D
PTF [397]. The other strategy is to exploit the illumination diversity and
OFS. For a single illumination aperture, the strong frequency response
and large spectral coverage of the PTF are difficult to be achieved simul-
taneously. Therefore, similar to the case of 2D QPI, we can incorporate
multiple intensity stacks with different illumination apertures to achieve
a synthesized 3D PTF, resulting in high-quality 3D RI reconstruction
with improved resolution and SNR. Note that unlike the multi-plane
TIE, the axial scanning is already included in the data acquisition, so
“multi-plane” is impossible and is meaningless for TIDT, leaving only
the illumination aperture is adjustable. So in order to obtain a more
complete frequency coverage and balanced transfer function, Li et al.
[214] proposed a multi-frequency combination scheme that incorpo-
rates three intensity stacks captured under three illumination apertures
including an annular aperture and two circular apertures with different
coherence parameters. The optimal frequency components of 3D PTFs
corresponding to multiple illumination apertures are combined together
by utilizing linear least-squares method, and a more accurate 3D recon-
struction result with imaging resolution up to the incoherent diffraction
limit can be obtained. Inspired by the resolution enhancement for 2D
QPI in Subsection 7.6, the annular illumination aperture is introduced
to not only achieve the maximum allowable lateral and axial resolu-
tion limit, but also maintain strong amplitude response of the 3D PTF,
preventing phase contrast vanishing under large circular illumination
aperture [Fig. 91(e)]. Fig. 92 gives the flowchart of the multi-aperture
TIDT based on multi-frequency combination. The Fourier spectra of the
corresponding three intensity stacks are taken as 𝐼̂1 (𝐮), 𝐼̂2 (𝐮) and 𝐼̂3 (𝐮)
under these three illumination apertures, and the corresponding PTFs
are HP1 (u), HP2 (u) and HP3 (u). The synthesized PTF can be expressed
as
( )
𝐼̂1 (𝐮) 𝐼̂ (𝐮) 𝐼̂ (𝐮)
𝑃 (𝐫 ) = ℱ −1
𝜀 + 2 𝜀 + 3 𝜀 (395)
𝐻 𝑃 1 (𝐮 ) 1 𝐻 𝑃 2 (𝐮 ) 2 𝐻 𝑃 3 (𝐮 ) 3
The weighting coefficients of the frequency component corresponding
to each pupil can be obtained by the following least squares method
𝐻 ∗ (𝐮)𝐻𝑃 𝑖 (𝐮)
𝑃𝑖
𝜀𝑖 = (396)
|𝐻 (𝐮)|2 + |𝐻 (𝐮)|2 + |𝐻 (𝐮)|2
| 𝑃1 | | 𝑃2 | | 𝑃3 |
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 91. 2D cross-sections of 3D PTFs for different (circular and annular) illumination apertures with different coherence parameters [214].

Fig. 92. Flowchart of the multi-aperture TIDT method based on multi-frequency combination [214].

Substituting Eq. (396) into the Eq. (395), the phase component can be 9. Applications of TIE in optical imaging and microscopy
finally reconstructed as follows
{̂ } In this section, we present optical configurations and representative
𝐼1 (𝐮)𝐻 ∗ (𝐮) + 𝐼̂2 (𝐮)𝐻 ∗ (𝐮) + 𝐼̂3 (𝐮)𝐻 ∗ (𝐮) applications of TIE with an emphasis on optical imaging, metrology, and
𝑃1 𝑃2 𝑃3
𝑃 (𝐫 ) = ℱ −1
(397)
|𝐻 (𝐮)|2 + |𝐻 (𝐮)|2 + |𝐻 (𝐮)|2 + 𝛼 microscopy. In Subsection 9.1, we present the conventional experimen-
| 𝑃1 | | 𝑃2 | | 𝑃3 |
tal setups for the TIE phase retrieval. In these setups, in order to ac-
Before closing this section, we need to further clarify the term “3D quire multiple intensity images at different defocus distances, either the
phase imaging”. It is generally known that the “phase” refers to the op- sample or the camera needs to be translated along the optical axis, ne-
tical path difference introduced by the non-uniform spatial distribution cessitating manual adjustment or mechanic devices. In Subsection 9.2,
of the RI or thickness of the object. It implies the meanings of “projec- we present several improved setups of TIE, which eliminate the me-
tion” and “accumulation”, which is defined on a 2D plane instead of in a chanical motion for through-focus intensity acquisition, extending TIE
3D volume. So the definition of “phase” is only valid for 2D thin object, to dynamic phase imaging applications. In Subsection 9.3, we present
otherwise it can only be used to represent the optical path integral of the optical configurations for TIE based diffraction tomography. Finally,
a 3D thick object along the light propagation direction. Consequently, in Subsection 9.4, we review the typical applications of TIE in the field
the term “3D phase imaging” is not rigorous. In this tutorial, the term of optical imaging and microscopy, and present some representative ex-
“3D phase imaging” is only adopted as a colloquial alternative to “3D perimental results.
RI diffraction tomography”.
C. Zuo, J. Li and J. Sun et al.
Fig. 93. Conventional optical configurations for the TIE phase imaging. (a) 4f system-based configuration; (b) lensless configuration.
9.1. Basic optical configurations for TIE phase imaging
The TIE phase retrieval method needs to collect intensity images at
multiple defocus planes. In order to capture these intensity images, it
is often necessary to use a 4f system, as illustrated in Fig. 93(a). Since
the object and image planes in the 4f system are strictly conjugated,
the intensity information at defocused planes can be obtained by either
moving the object or the camera. These two methods are essentially
equivalent, but when considering lateral and axial magnifications, the
object-space defocus distance is scaled by a factor of 𝑓22 ∕𝑓12 when the de-
focusing is achieved by moving the camera in the image space. For an
optical microscope with an infinity-corrected optical system, the optical
path is equivalent to Fig. 93(a). The collimator of Fig. 93(a) is equivalent
to the condenser lens in the Köhler illumination, while the lenses L1 and
L2 correspond to objective and tube lens in the infinity-corrected micro-
scope, and f2 /f1 is the magnification of the microscope. Therefore, an
infinity-corrected microscope can be directly used for TIE retrieval with-
out any modifications. Besides, TIE can also adopt the lensless imaging
setup shown in Fig. 93(b), in which the intensity images are all taken at
the out-of-focus plane. So the phase reconstructed by TIE is also located
at the defocused plane, and additional numerical back-propagation or
digital refocusing algorithms are required to reconstruct the complex
amplitude distribution of the sample at the object plane.
9.2. Advanced optical configurations for dynamic TIE phase imaging
In traditional optical setups of TIE, one needs to move the object or
camera manually or mechanically when capturing the intensity images,
which inevitably reduces the speed of data acquisition, precluding dy-
namic, high-speed, and real-time measurement applications. To address
this problem, numerous advanced optical configurations for dynamic
TIE phase imaging have been proposed in recent years. Their common
feature is to avoid mechanical motion during the image acquisition pro-
cess. For example, Blanchard et al. [398] placed a quadratically distorted
grating in front of the imaging lens, creating three laterally shifted im-
ages corresponding to different defocus distances in a single image plane
[Fig. 94(a)]. Waller et al. [191] utilized the chromatic aberration inher-
ent in the microscope to obtain three intensity images at different axial
distances from a single color image. Waller et al. [190] also proposed
single-shot TIE optical configurations based on a multiplexed volume
hologram to laterally separate images from different focal planes, en-
abling acquisition of intensity images at different depths from a single
exposure [Fig. 94(b)]. Almoro et al. [193] realized non-mechanical im-
age defocusing by introducing a SLM in the Fourier plane of the 4f imag-
ing system [Fig. 94(c)]. Gorthi and Schonbrun [194] applied TIE to flow
cytometry and realized the automatic collection of through-focus inten-
sity images by tilting the microfluidic channel vertically [Fig. 94(d)].
Martino et al. [399] developed a single-shot TIE setup based on a beam
splitter and several mirrors, permitting recording two intensity images
at different depth simultaneously [Fig. 94(e)].
Optics and Lasers in Engineering 135 (2020) 106187
Not just limited to independent imaging systems, TIE is also fully
compatible with an off-the-shelf infinity-corrected microscope. Dynamic
QPI can also be achieved by introducing additional optics to the camera
port of the microscope. Zuo et al. [216,217] proposed two microscope-
based dynamic TIE phase imaging systems. The tunable lens based
TIE (TL-TIE) system introduces an electrically tunable lens (ETL) and
a 4f system into a conventional bright-field microscope. The TL-TIE
permits high-speed axial scanning to realize the non-mechanical focus
control with constant magnification for high-speed TIE phase imaging
[217] [Fig. 95(a)]. The single-shot quantitative phase microscopy (SQPM)
system realizes the programmable digital defocusing based on an SLM.
Based on a Michelson-like optical architecture attached to an unmod-
ified inverted bright-field microscope, two laterally separated images
from different depths can be obtained simultaneously by a single cam-
era exposure, enabling TIE phase imaging at the camera limited speed
[216] [Fig. 95(b)].
9.3. Optical configurations for TIE phase tomography
Although TIE has been successfully used for phase retrieval and QPI,
the fact that phase is a product of thickness and average RI can po-
tentially result in inaccurate interpretations of QPI data when imaging
complex 3D structures. In order to gain more accurate morphological
information, diffraction tomography is often desired. Compared with
TIE phase imaging, the optical configurations for diffraction tomogra-
phy based on TIE are relatively complicated, and there are two typical
solutions. The first category of techniques is based on traditional CT or
ODT models, which obtains multiple quantitative phase distributions by
sample rotating or beam scanning. Different from traditional diffraction
tomography based on interferometry or digital holography, the com-
plex amplitude of the field under each illumination angle is retrieved
by TIE in a noninterferometic manner. In 2000, Barty et al. [175] com-
bined TIE phase retrieval with sample rotation based on a traditional
microscope platform. Two different optical fibers were fixed with the
tip of a syringe under the control of a rotating motor [Fig. 96(a)]. The
RI profiles of the fibers were reconstructed by the inverse Radon trans-
form. In 2011, based on similar ideas and experimental setups, Lee et al.
[400] achieved 3D tomography of a glass diamond fixed on a rotat-
ing stage based on the back filtered projection method. Subsequently,
Nguyen et al. [401,402] measured the 3D RI distributions of fibers and
gemstone microspheres fixed on a rotating turret based on the TL-TIE
[217] and SQPM systems [216]. In those early works, the 3D recon-
struction physical model is similar to what was used in x-ray CT, which
assumed phase measurement to be an integration of RI along the projec-
tion angle and ignored the significance of the optical diffraction effect,
thus limiting the earlier TPM systems for objects with simple structures.
In biology imaging, this inaccurate tomography model can significantly
affect the 3D reconstruction resolution, especially for cells that are much
thicker than the depth of field of the imaging systems. In 2015, Zuo et al.
[212] proposed a lensless diffraction tomographic platform using only
an image sensor and a programmable color LED array. By illuminating
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 94. Several improved optical configurations for dynamic TIE phase imaging. (a) optical configuration based on quadratically distorted grating [398]; (b) optical
configuration based on volume holography [190]; (c) optical configuration based on spatial light modulator [193]; (d) optical configuration based on flow cytometry
[194]; (e) optical configuration based on beam splitter and multiple reflections [399].

Fig. 95. Two optical configurations for dynamic TIE phase imaging, which can be implemented as a simple add-on module to a conventional off-the-shelf microscope.
(a) electrically tunable lens based TIE microscopic system (TL-TIE) [217]; (b) single-shot TIE system based on a SLM (SQPM) [216].
C. Zuo, J. Li and J. Sun et al.
Fig. 96. Optical configurations for 3D diffraction tomography based on TIE. (a) Tomographic system in which the sample is fixed by a needle tip and rotated by a
motorized stage [175]; (b) TIE diffraction tomographic system based on multi-angle illuminations with an LED array [213]; (c) TIE diffraction tomographic system
based on intensity axial scanning with an off-the-shelf microscope [214].
the sample at multiple wavelengths, the quantitative phase of the spec-
imen can be retrieved by TIE. Based on the wavelength dependence of
Fresnel propagation, this approach treats the diffraction patterns gener-
ated by RGB three-color illuminations as intensity images at three dif-
ferent defocus distances. Furthermore, by illuminating the sample se-
quentially with different LEDs across the full array, the complex fields
of the object at different illumination angles can be combined into the
3D RI distribution based on the ODT model under the first Rytov approx-
imation. In 2015, Jenkins et al. [380] combined the object rotation and
axial scanning to realize 3D tomographic reconstruction of the RI distri-
butions of different types of optical fibers. In 2017, Li et al. [213] pro-
posed a TIE tomography system based on an off-the-shelf microscope by
replacing the Köhler illumination system of the traditional bright-field
microscope with a programmable LED array [Fig. 96(b)]. The intensity
images of the sample are captured in two different defocused planes at
each illumination angle, and the corresponding phase distributions are
recovered by TIE. Finally, the 3D RI information of the object can be
reconstructed based on the Fourier diffraction theorem. Different from
the interferometric tomographic system where the incident laser illumi-
nation is scanned with the help of galvanometer mirrors, the LED-array-
based configurations are inherently much simpler and achieve improved
imaging stability and speed.
Since 2017, TIDT has gradually emerged as a new diffraction to-
mography technique and attracted much attention of researchers. TIDT
is a partially coherent ODT approach, which only needs to capture the
through-focus intensity images across the object, and then uses the im-
age deconvolution algorithm to directly retrieve the 3D RI distribution
Optics and Lasers in Engineering 135 (2020) 106187
of the object. Thus, it effectively bypasses the difficulties associated with
the traditional interferometric diffraction tomography techniques, such
as interferometric measurements, object rotation, and beam scanning.
The hardware setup is fully compatible with conventional bright-field
microscope systems. Soto et al. [381] realized the TIDT based on the TL-
TIE system [217]. An electrically tunable lens is employed to axially scan
the sample in high speed. The resultant through-focus stack is deconvo-
luted based on 3D WOTF to recover the 3D RI distribution of the sam-
ple. In 2018, Li et al. [214] combined TIDT with illumination coherence
engineering. Three sets of through-focus intensity images are captured
under an annular aperture and two circular apertures with different co-
herence parameters [Fig. 96(c)]. By changing the condenser aperture
diaphragms of a traditional bright-field microscope, the 3D WOTFs cor-
responding to different illumination apertures are combined to obtain
a synthesized frequency response, achieving high-quality, low-noise 3D
reconstructions with imaging resolution up to the incoherent diffraction
limit.
9.4. Applications of TIE in optical imaging and microscopy
Similar to many other phase retrieval algorithms, TIE was mainly
oriented to adaptive optics [144–148], X-ray diffraction imaging
[161,162], STEM [166,169,171,172], neutron radiography [164,165],
transmission electron microscopy etc.. In recent years, as the TIE tech-
nique advances and matures, it has been increasingly applied to the op-
tical imaging and microscopy of the visible light band. Due to the space
C. Zuo, J. Li and J. Sun et al.
Fig. 97. TIE phase-contrast imaging of testicle of a rat [142]. (a) bright-field
intensity image; (b) axial intensity derivative distribution.
Fig. 98. Quantitative phase imaging of an unstained cheek cell [175]. The insets
show the DIC image of the cell and the recovered quantitative phase image by
TIE.
limitation, this section reviews several representative applications of
TIE, focus particularly on optical imaging, metrology, and microscopy.
9.4.1. QPI for bio-applications
The applications of TIE to optical microscopy can be traced back
to 1984 (only about one year after Teague deduced TIE). Streibl
[142] briefly proved that TIE is applicable to phase-contrast imaging
of transparent phase objects under partially coherent Köhler illumina-
tion of a off-the-shelf microscope. He demonstrated the phase gradient
enhancement effect of unstained mouse testis cells (Fig. 97). Since the
numerical solution to TIE has not been proposed at that time, he only
presented the axial intensity derivative image rather than the quanti-
tative phase reconstruction result. Streibl believed that the axial inten-
sity derivative image could effectively highlight phase details that are
difficult to be observed in the bright-field image. In 1998, Barty et al.
[175] briefly reported the QPI results of cheek cells and optical fibers
based on TIE (Fig. 98). Strictly speaking, it was the debut of TIE in
the field of quantitative phase microscopy. They also emphasized that
the phase recovered by TIE is continuous, and no phase unwrapping
is needed. The beautiful experimental results presented by Barty et al.
has opened the door for the subsequent applications of TIE in biologi-
cal and biomedical imaging. In 2004, Curl et al. [180,181] used TIE to
quantitatively detect the morphology and growth rate of cells, indicating
that quantitative phase information is very conducive to data process-
ing and quantitative analysis, such as cell segmentation and counting.
In 2005, Ross et al. [182] utilized TIE to enhance the imaging contrast of
unstained cells under microbeam irradiations. In 2007, Dragomir et al.
[183] successfully applied TIE to the quantitative measurement of the
birefringence effect of unstained isolated cardiomyocytes.
In 2010, Kou et al. [188] demonstrated that TIE can be directly imple-
mented based on an off-the-shelf DIC microscope to realize QPI, and suc-
cessfully obtained the quantitative phase images of human cheek cells.
Optics and Lasers in Engineering 135 (2020) 106187
Fig. 99. Quantitative phase imaging of an unstained cheek cell [192]. The insets
show the phase-gradient image from quantitative phase reconstructed by TIE
and phase-shifting DIC, respectively.
Fig. 100. Quantitative phase imaging and RBC screening [403]. (a) Scatterplot
of RBCs subjected to isotonic and hypotonic solutions; (b)-(i) phase images in
the gated regions marked in (a); Normal RBCs (b)-(e) in flow took parachute like
or slipper like shapes; Spherized RBCs (f)-(i) remained spherical even in flow.
The phase reconstruction results were in good agreement with those
obtained by the polarization phase-shifting technique (Fig. 99). Waller
et al. proposed single-shot TIE optical configurations based on volume
holography [190] and chromatic aberration [191], enabling acquisition
of intensity images at different defocus distances from a single exposure.
Based on these setups, they presented the QPI results of deformable mi-
cromirror arrays and HeLa cells. In 2011, Kou et al. [192] put forward
the WOTF deconvolution method for QPI under partially coherent il-
luminations and quantitatively compared it with the conventional TIE.
The QPI results of human Ascaris are very similar to those obtained by
TIE, which is understandable because the low-frequency component of
the WOTF under weak defocusing is a Laplacian in the frequency do-
main (TIE’s PTF). In 2012, Gorthi and Schonbrun [194] first applied
TIE to flow cytometry and realized the automatic collection of through-
focus intensity images by tilting the microfluidic channel vertically. The
high-throughput sorting of red blood cells was demonstrated based on
morphological information extracted from their quantitative phase im-
ages. In the same year, Phillips et al. [404] measured the RI, dry mass,
volume, and density of red blood cells based on TIE with use of a tradi-
tional bright-field microscope (though they claimed that they obtained
3D RI distribution of the cell, their multi-plane TIE tomography method
is not rigorous in principle). Subsequently, Phillips et al. [403] further
applied their method for circulating tumor cell (CTC) screening in an ovar-
ian cancer patient: the TIE phase retrieval was employed to measure
cellular dry mass, and the reconstructed DIC images were used to mea-
sure cellular volume. They found that high-definition CTCs were more
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Fig. 101. Quantitative phase imaging of an in-

dividual MCF-7 cell based on TL-TIE [217]. (a-
c) Intensity images with defocus distance -2.5,
0, 2.5 𝜇m; (d) phase map recovered by TIE; (e)
digitally simulated DIC image from (d); (f) 3D
pseudo-color rendering of the cell thickness.

(a) (b) (c)

Fig. 103. Delineation of HeLa cells [275]. (a) In-focus bright-field image; (b)
phase image obtained by using TIE reconstruction; and (c) segmentation result
with watershed algorithm.

Fig. 102. Dynamic TIE phase imaging of macrophage phagocytosis [216]. (a)
Color-coded phase profiles at different stages of phagocytosis; (b) phase maps
of the nuclear region of the macrophage-the black square in (a)-during the in-
ternalization stage of phagocytosis; (c) phase/thickness variation with time of
three points, indicated by the dots in (b): (A) red, (B) green, and (C) blue, and
the average of the whole square region (bottom, black curve). Scale bar: 10 𝜇m.

massive than leukocytes, which provides key insights into the fluid dy-
namics of cancer, and may provide the rationale for strategies to isolate,
monitor or target CTCs based on their physical properties.
In 2013, Zuo et al. [217] demonstrated the dynamic QPI results of
a live MCF-7 breast cancer cell based on the TL-TIE system. The high-
resolution quantitative phase images revealed the subcellular dynam-
ics, i.e., the nucleus inside the cell, ruffles on the membrane, and the
lamellipodia in the cell periphery (Fig. 101). In the same year, Zuo
et al. [216] investigated of chemical-induced apoptosis and the dy- Fig. 104. Multi-modal imaging of live HeLa cells [406]. (a) Quantitative phase
namic phagocytosis process of RAW264.7 macrophages (Fig. 102) based map of the whole FOV. (b-c) experimental results for bright-field intensity,
on SQPM. The high-resolution dynamic 3D QPI videos of macrophage phase contrast, DIC, and quantitative phase images of mitotic cell and interphase
phagocytosis were presented, suggesting TIE is a simple and promising cells.
approach for label-free imaging and quantitative study of cell dynamics
at subcellular resolution. In 2016, Bostan et al. [275] used TIE to re-
cover the quantitative phase images of HeLa cells. The measured phases tal ZPC, DIC, QPI, and light-field imaging of HeLa cells under division
were consistent with those obtained by digital holographic microscopy. (Fig. 104). It should be mentioned that a similar multi-modal TIE imag-
They also pointed out that quantitative phase images are extremely con- ing approach was previously presented by Paganin et al. [405] in lensless
venient for subsequent cell segmentation and counting (Fig. 103). In X-ray diffraction imaging.
2016, Li et al. [215] realized multi-modal microscopic imaging based In 2017, Zuo et al. [209] employed the annular illumination based
on TIE based on an off-the-shelf bright-field microscope, realizing digi- TIE for long-term time-lapse imaging of HeLa cell dividing in culture.
C. Zuo, J. Li and J. Sun et al.
Fig. 105. Time-lapse TIE phase imaging of HeLa cell division over a long pe-
riod (60 h) [209]. (a) Representative quantitative phase images at different time
points; (b) the change of cell number and confluence ratio over the culture pas-
sage period; (c) and (d) show the magnified views corresponding to two regions
of interest (Area 2 and Area 3); (e) 10 selected time-lapse phase images and the
corresponding 3D renderings showing the morphological features of a dividing
cell (Area 1) at different stages of mitosis.
Fig. 106. Dual-mode confocal fluorescence and TIE phase imaging in a confocal
laser scanning microscope [407]. (a) Schematic depiction of the optical setup;
(b) exemplary fluorescence; and (c) optical path difference images measured
with this setup. Scale bar, 20𝜇m; color bar labels, nm.
The high-resolution quantitative phase images allow for assessing the
progression of confluence ratio over 60 hours (Fig. 105). Zheng et al.
[408,409] introduced TIE into Coherent Anti-Stokes Raman Scattering
(CARS) microscopy. The Several CARS images at different defocused
planes perpendicular to the propagation direction were recorded to re-
construct a quantitative phase map based on TIE, which allows for non-
resonant background suppression in CARS microscopy. In 2018, Li et al.
[406] realized dynamic TIE phase imaging of live human osteoblasts
based on a modified SQPM system. Liao et al. [410] combined the TIE
phase retrieval technique with a self-developed dual-LED-based auto-
focusing microscope, achieving the whole slide QPI of an unstained
mouse kidney section. Zheng et al. [407] proposed the dual-mode phase
and fluorescence imaging in a confocal laser scanning microscope by
combining the confocal microscopy with TIE phase imaging (Fig. 106).
The limited depth of field was extended about sixfold by axial scan-
ning with a tunable acoustic gradient index of refraction lens, making
TIE compatible with confocal microscopy. The dual-mode images with
Optics and Lasers in Engineering 135 (2020) 106187
pixel-to-pixel correspondence enable the standalone determination of
the RI of live cells.
In 2019, Li et al. [411] quantitatively characterized the cell morphol-
ogy and cytoskeletal dynamics of living osteoblastic cells based on TIE.
The changes in mass transport are visualized based on phase correlation
imaging [412]. Bian et al. [413] proposed a miniaturized portable micro-
scope to assess the inhomogeneous optical properties of tissues based on
the quantitative phase maps obtained by TIE. Rajput et al. [414] demon-
strated that the quantitative phase recovered by TIE could be used for
nonscanning 3D fluorescence imaging. Though the propagation of inco-
herent fields does not obey the Fresnel diffraction law, and the phase
of an incoherent field is not well-defined, their experimental results on
neurite and cells did show certain feasibility. In 2020, Mandula et al.
[415] combined TIE with multi-color fluorescence for long-term dual-
mode imaging of dividing HeLa cells based on a standard bright-field
microscope platform. The changes of the focus are achieved with specif-
ically introduced chromatic aberration in the imaging system, allow-
ing motion-free image acquisition. Wittkopp et al. [416] compared the
QPI results of resolution target and check cells obtained by DHM and
TIE based on a custom-built microscope system, revealing quantitative
agreement between the two methods in both static and live samples.
9.4.2. Phase tomography for bio-applications
The key difference between phase tomography and QPI is the ability
to obtain the “true 3D” RI profile of a biological sample. In 2015, Zuo
et al. [212] demonstrated the lensless quantitative phase microscopy and
diffraction tomographic imaging of fertilized eggs of Parascaris equo-
rum based a lensless compact on-chip platform with a programmable
color LED array (Fig. 108). By illuminating the sample with a multi-
wavelength LED, the quantitative phase of the specimen can be retrieved
by TIE. With the quantitative phase retrieved, the diffracted object field
can be digitally refocused through back propagation without using any
lenses. Furthermore, by illuminating the sample sequentially with dif-
ferent LEDs across the full array, the complex fields of the same object
from different illumination angles can be obtained. Finally, the 3D RI
distribution of the sample is reconstructed based on the ODT model.
The miniaturized system offers high quality, depth-resolved images with
high resolution (lateral resolution 3.72𝜇m, axial resolution 5𝜇m) and
across a wide FOV (24mm2 ). In 2017, Soto et al. [381] reconstructed
the 3D RI distribution of human RBCs based on TIDT [Fig. 107(a)-(c)].
This technique was subsequently applied to 3D quantitative RI recon-
struction of dynamic biological samples by Rodrigo et al. [417]. They
demonstrated 3D RI visualization of living bacteria both freely swim-
ming and optically manipulated by using freestyle laser traps allowing
for their trapping and transport along 3D trajectories. In 2017, Li et al.
[214] demonstrated quantitative RI imaging based on TIDT under mul-
tiple illumination apertures. Experimental results verified that the use of
multiple illumination apertures provides a broader frequency coverage
and higher frequency responses. Fig. 107(d)-(e) show 3D RI rendering
results of a unstained Pandorina morum algae and a HeLa cell recon-
structed by Li et al. [214]. In 2019, Ma et al. [418] presented a simpli-
fied TIDT approach based on spatially-incoherent annular illumination
for weakly absorbing samples, e.g., tissues and diatoms. A 3D gradi-
ent operation was adopted to remove the background, leaving only the
3D skeleton structure of the weakly absorbing sample visible in the im-
ages. Li et al. [419] proposed a high-speed in vitro intensity diffraction
tomographic technique utilizing annular illumination to rapidly char-
acterize large-volume 3D RI distributions with native bacteria, and live
Caenorhabditis elegans specimens. Soto et al. [420] reported the use of
TIDT for quantitative analysis of unstained DH82 cell line infected with
Leishmania infantum. The experimental results demonstrated that the in-
fected DH82 cells exhibit a higher RI than healthy samples.
9.4.3. Optical metrology
As a quantitative phase measurement technique in nature, TIE was
also widely used in the field of optical metrology. As early as in 1998,
C. Zuo, J. Li and J. Sun et al.
Fig. 107. 3D Tomographic reconstruction of several biological samples. (a)-(c) bright-field image, RI slices, and 3D rendering of RI distribution of human blood cell
sample [381]; (d)-(e) 3D rendering of RI tomograms of a Pandorina morum algae and a HeLa cell [214].
Fig. 108. 3D tomographic imaging of a slice of the uterus of Parascaris equorum
[212]. (a) RI; (b) absorption sections at different depth. The second row shows
the corresponding 𝑥 − 𝑧 views of 3D stacks (scale ratio of z to x axis is 1:2). The
arrows point out the dust particle located at a higher layer (at z = 168.4𝜇m); (c)
3D rendering of the RI; (d) 3D rendering of the absorption distribution. Scale
bar: 400𝜇m.
Barty et al. [175] successfully applied TIE to characterize the phase pro-
files of optical fibers immersed in RI-matching medium in a conventional
optical microscope platform. In 2000, Barty et al. [176] extended this
Optics and Lasers in Engineering 135 (2020) 106187
Fig. 109. Quantitative phase tomography of a single-mode fiber [176]. (a) 2D
slice through the 3D reconstruction perpendicular to the rotation axis; (b) line
profile through the middle of the reconstructed 3D distribution (dashed line)
compared to the known RI distribution (solid line).
method to phase tomography. They measured the phase distribution of
the sample at various angles by rotating the object, and 3D RI distri-
bution of various optical fibers (single-mode, multi-mode) can be syn-
thesized according to the inverse Radon transform (Fig. 109). In 2002,
based on the axisymmetric assumption of fiber, Roberts et al. [184] ob-
tained the radial RI distribution of the fiber only by preforming inverse
Abel transform on the phase distribution measured by TIE. The exper-
imental results are consistent with the exponential profile obtained by
the commercial profiler, and the relative error is lower than 0.0005. Due
to the simplicity and effectiveness, this method has also attracted exten-
sive attention and been further improved [185,187,421]. In 2007, Dor-
rer and Zuegel [186] used TIE to quantitatively characterize the topo-
logical variations induced by the magnetorheological finishing process
on laser rods. They successfully applied the technique to the Omega EP
laser device (OMEGA EP Laser Facility). In 2010, Almoro et al. [69] used
C. Zuo, J. Li and J. Sun et al.
Fig. 110. Characterization of a plano-convex quartz microlens array (pitch
250𝜇m) [202]. (a) In-focus intensity; (b) defocused intensity distribution (Δ𝑧 =
−550𝜇𝑚); (c) axial intensity derivative, the inset shows the enlarged boxed re-
gion; (d) retrieved phase; (e) Rendered surface plot; (f) confocal microscopic
result; (g) 3D topography by confocal microscopy; (f) Comparison of the line
profiles of single lens.
speckle field illumination to enhance the phase contrast of the defocused
intensity images for smooth objects. They demonstrated that this tech-
nique could effectively improve the phase measurement accuracy for
lens characterization. In the same year, Shomali et al. [422] proposed the
idea of applying TIE to aspheric optics testing. The traditional method is
based on “null test”, e.g., by using a wavefront-matching diffractive op-
tical element (DOE) or tilted wavefront interferometry, resulting in ex-
tremely complex and expensive optical configurations. TIE only needs
to capture several intensity images of the sample at different defocus
distances to recover the absolute phase of the complex surface directly.
It effectively bypasses the phase aliasing problem arising from the over-
dense fringes around large wavefront curvature in traditional interfer-
ometry.
In 2014, Zuo et al. [202] also extended TIE to the characterization
of micro-optical components. With use of the DCT-based TIE solver, dif-
ferent types of micro-optical components, e.g., microlens arrays, cylin-
drical microlenses, Fresnel lenses were accurately characterized with-
out any boundary artifacts which is common for FFT-based solver (the
periodic boundary condition is not satisfied). Fig. 110 shows the mea-
surement results for a plano-convex quartz microlens array with a pitch
of 250 𝜇m. It can be seen that the curvature radius recovered by the TIE
method is 346.7 𝜇m, and the result obtained by confocal microscopy is
350.4 𝜇m, both these two results are consistent with the standard ref-
erence value (350 𝜇m) provided by the manufacturer. In 2015, Jenkins
et al. [380] combined the object rotation and axial scanning to real-
ize 3D tomographic reconstruction of the RI distributions of different
types of optical fibers, including single-mode, polarization-maintaining,
and photonic-crystal fibers as well as an azimuthally varying CO2-laser-
induced long-period fiber grating period as test phase objects. In 2017,
Pan et al. [423] reconstructed the complex amplitude distribution of the
laser beam based on TIE, and then used the angular spectrum propa-
gation method to obtain the beam intensity distribution of the optical
field in the arbitrary cross-section along the propagation direction, and
finally calculated the M2 quality factor of the beam. Experimental re-
sults of He-Ne and high-power fiber laser sources indicated that this
method could obtain a more accurate M2 quality factor, and the mea-
sured values coincide with the results obtained by the beam propagation
analyzer. In 2019, Bunsen and Tateyama [424] used TIE to detect the
phase distribution of the signal beam of holographic data storage (HDS),
Optics and Lasers in Engineering 135 (2020) 106187
Fig. 111. Lensless TIE phase imaging system with multi-wavelength LED illu-
minations [212]. (a) Schematics explaining the principle of lensless imaging; (b)
photograph of the system.
contributing to miniaturization of the optical system and improvement
of the vibration tolerance of HDS.
9.4.4. Lensless imaging
On-chip lensless holographic microscopy is a new computational
imaging technology in which the sample is placed directly on the the
surface of the image sensor, and the resultant diffraction pattern un-
der quasi-coherent illumination is recorded. The image is reconstructed
by numerical back propagation once the phase of the diffraction field
can be retrieved. Lensfree holographic microscopy was originally pro-
posed in 2009 by the Ozcan group of UCLA [425]. Compared with con-
ventional lens-based microscopic systems, the lensless imaging configu-
ration enjoys two distinct advantages: first, the imaging magnification
is close to 1 due to the small distance between the object and the de-
tector (generally less than 1mm), i.e., the imaging FOV is completely
determined by the size of the photosensitive area of the image sensor;
second, the effective NA of the imaging system is close to 1 due to the
small sample-to-sensor distance, i.e., almost forward scattering light can
be acquired by the image sensor. Initially, the lensless on-chip micro-
scopic technique is mostly based on iterative phase retrieval algorithms,
which recover the phase distribution based on the assumption of sparse
samples [70,425] or multi-distance intensity measurements [426,427],
and then the intensity and phase distribution at the object plane can
be reconstructed by the angular spectrum back propagation algorithm.
In 2015, Zuo et al. [212] developed a miniaturized lensless TIE tomo-
graphic microscope based on a programable LED array (Fig. 111). With
the interchangeability of illumination wavelength and propagation dis-
tance in the Fresnel domain, the quantitative phase of the specimen
can be retrieved through TIE with multi-wavelegth LED illuminations.
By illuminating the sample sequentially with different LEDs across the
full array, the complex fields of same object from different illumination
angles can be mapped in 3D Fourier space, and the tomogram of the
sample can be reconstructed. This was the first demonstration of TIE for
lensless phase microscopy and diffraction tomography.
After TIE techniques gained considerable attention in microscopic
imaging, it was also gradually introduced into the field of lensless mi-
croscopy. Since only 2-3 intensity images are sufficient to recover the
phase distribution, TIE can be used to create an initial estimate of the
phase distribution quickly and conveniently [212,429,433]. The initial
estimate is then optimized by iterative phase retrieval to improve its spa-
tial resolution and compensate for the phase discrepancy due to parax-
ial approximation and Teague’s assumption. On the other hand, lens-
less holographic microscopy shifts the problem of the limited SBP of
conventional microscopic systems to the limited sampling spacing of
image sensors. In principle, the key factor limiting the imaging reso-
lution of lensless microscopy is the high-frequency signal aliasing in-
C. Zuo, J. Li and J. Sun et al.
Fig. 112. Schematic diagrams of the pixel super-resolution based lensfree on-chip imaging setup. (a) Sub-pixel shifting of illumination source [428]; (b) 2D horizontal
sub-pixel sensor motion [429]; (c) fiber-optic array based source scanning [430]; (d) illumination wavelength scanning [431]; (e) axial scanning with multiple sample-
to-sensor distances [426]; (f) active source micro-scanning using parallel plates [432].
duced by undersampling, rather than the loss in high-frequency de-
tails induced by the limited aperture [434]. Therefore, a lot of research
works in this field focused on “pixel super-resolution”, which reduces
the effective pixel size through computational approaches so as to re-
alize “sub-pixel imaging”. Representative methods include: sub-pixel
shifting of illumination sources [428] [Fig. 112(a)], sub-pixel lateral
motion of image sensor [429] [Fig. 112(b)], fiber-optic array-based
source scanning [430] [Fig. 112(c)], illumination wavelength scanning
[431] [Fig. 112(d)], axial scanning with multiple sample-to-sensor dis-
tances [426] [Fig. 112(e)], etc.. In 2018, Zhang et al. [432] proposed
an active sub-pixel micro-scanning scheme based on two parallel plates
[Fig. 112(f)], which realized high-precision controllable sub-pixel scan-
ning with low-cost mechanical components. By combining TIE and GS
phase retrieval methods, the Nyquist-Shannon sampling resolution limit
imposed by the pixel size of the imaging sensor was effectively sur-
passed, and the high-resolution wide-FOV dynamic QPI can be achieved
(effective imaging NA ≈ 0.5).
9.4.5. Light-field imaging
In Section 7, we discussed that conditional frequency moment (cen-
troid) of the light field could be recovered by TIE. The moment repre-
sents the average direction of the light rays passing through each spatial
point. In 2013, Orth and Crozier [364] reported a computational light-
field imaging method termed LMI to extract the first angular moment of
light field. They derived a continuity equation to reconstruct the light
field with only a pair of intensity images exhibiting a slight defocusing.
Optics and Lasers in Engineering 135 (2020) 106187
The “moment” of “light field moment imaging” means that this method
can only recover the first angular moment of the light field rather than
the complete 4D light field information. In order to “fill in” the miss-
ing data, Orth and Crozier [364] assumed that the angular distribution
of the light field is Gaussian. Although this assumption lacks physical
evidence, their experimental results indeed exhibit good visual effects.
In 2014, Zuo et al. [207] commented that LMI could be associated with
TIE under the geometric-optics limit, so all numerical solutions and ax-
ial intensity derivative estimation algorithms in TIE can be directly ap-
plied to LMI. For example, Liu et al. [365] employed the multi-plane
high-order finite-difference method to improve the accuracy of axial in-
tensity derivative estimation, thus reduced the noise and artifacts in the
reconstructed light field. In 2015, Zuo et al. [208] derived the GTIE for
partially coherent fields in phase space based on the transport equation
of WDF. Under the geometric-optics limit, WDF is equivalent to light
field [335], and TIE can be associated with the light-field imaging. Zuo
et al. [208] deduced that under certain conditions (e.g., slow-varying
objects under spatially stationary illuminations), the 4D light field is
highly redundant, and the complete 4D light field can be fully recovered
by TIE with the pre-knowledge of the source intensity distribution. In
2018, Wang et al. [435] presented an efficient data acquisition approach
for LMI with precise focal plane sweeping controlled by a SLM rather
than mechanical movement. Chen et al. [436] presented an overview
of 3D imaging from depth measurements, and provided a summary of
the connections between the ray-based and wavefront-based 3D imaging
techniques.
C. Zuo, J. Li and J. Sun et al. Optics and Lasers in Engineering 135 (2020) 106187

Table 8
Comparison of digital holography and transport of intensity equation

Terms Digital holography Transport of intensity equation

Data formulation Interferometric Non-interferometric

Beam coherence requirement
Raw data type
Numerical processing method
Main noise type
Resolution limit
Dynamic measurement
Phase recovered range
High
Hologram (interferogram)
Phase demodulation and numerical propagation
Spatial coherent noise (speckle)
CCD size and pixel resolution (lensless geometry),
coherent diffraction limit (lens-based geometry)
Off-axis geometry
Wrapped to [−𝜋, 𝜋) (need phase unwrapping)
Low, can be partially coherent
Through-focus intensity images (diffraction patterns)
Solving partially differential equation
Low-frequency noise (cloud-like artifacts)
Diffraction limit of partially coherent imaging
(illumination aperture, numerical aperture)
Need specific configurations
Continuous (no unwrapping required)

9.4.6. Other applications
In addition to the above-mentioned imaging and measurement ap-
plications, TIE was recently extended to become a spatial phase unwrap-
ping algorithm. Since TIE can directly obtain a continuous phase with-
out phase wrapping, it seems that it can effectively bypass the phase
unwrapping step, as is common for interferometric phase measurement
methods. Based on this characteristic, in 2013, Zuo et al. [277] intro-
duced TIE into digital holography. Three intensity images with slightly
defocusing were numerically calculated based on Fresnel diffraction
from the complex amplitude reconstructed by digital holography. By
solving TIE with artifactually created intensity images, the continuous
phase can be recovered. Zuo et al. [277] also pointed out that this
method can also be used as a spatial phase unwrapping algorithm by
assuming a uniform intensity distribution. It should be noted that no
phase unwrapping is a major advantage of the TIE phase retrieval, but
it assumes that the phase is a single-valued continuous function, which
does not fundamentally address the 2𝜋 ambiguity problem in traditional
phase measurement techniques. In 2016, Pandey et al. [437] proposed
a similar phase unwrapping method based on TIE and demonstrated
that this method is superior to the traditional path-dependent Goldstein
method. Subsequently, Zuo et al. [438] pointed out that TIE under uni-
form intensity is a Poisson equation, and TIE phase unwrapping is es-
sentially equivalent to the standard least-squares phase unwrapping al-
gorithm. In 2017, Martinez-Carranza et al. [439] further improved TIE
phase unwrapping algorithm. The idea of the least-squares phase un-
wrapping algorithm was used to calculate the Laplacian of the wrapped
phase efficiently, the time-consuming computation associated with the
wave propagation can be avoided, and the phase reconstruction accu-
racy can be effectively improved. In 2018, Zhou et al. [440] experi-
mentally demonstrated that the continuous phase demodulated from a
digital hologram by TIE is less noisy than the one obtained by conven-
tional holographic reconstruction. In 2019, Cheng et al. [441] proposed
a dual-wavelength phase unwrapping algorithm based on TIE. Zhao et al.
[442] used the iterative DCT TIE solver to correct the phase unwrapping
error iteratively, which improves the robustness of the TIE phase un-
wrapping under noisy conditions. In addition to phase unwrapping, TIE
has also been applied to optical encryption [443–446], and microscopic
imaging in miniaturized mobile phone platforms [447,448].
10. Conclusions and future directions
The development and fulfillment of TIE marks an important step for
phase retrieval and QPI technologies: the transition from interferomet-
ric to non-interferometric, from iterative to deterministic, from phase
recovery to light-field imaging, and from completely coherent to par-
tially coherent or even incoherent. This tutorial has reviewed the cur-
rent status and latest developments of TIE, including its basic principles,
numerical solutions, axial intensity derivative estimation, partially co-
herent imaging, 3D diffraction tomography, and its representative ap-
plications in optical imaging, metrology, and microscopy. TIE provides
an advanced and invaluable label-free tool for biological imaging and
biomedical research. It does not require strictly spatially and temporally
coherent illuminations, making it fully compatible with any off-the-shelf
bright-field microscopes. Thanks to the built-in Köhler illumination op-
tics and aberration-corrected objective lens, high-quality quantitative
phase images with diffraction-limited resolution can be easily obtained
with no worries about the parasitic interference and speckle noise. It
can be easily combined with fluorescence techniques to gain molecular
specificity and thus provides a wider window to investigate biological
processes. In addition, its single-beam (common-path) configuration is
highly stable, and therefore this technique is inherently insensitive to
environmental vibrations and other disturbances, eliminating the need
for bulky vibration isolation platforms. Furthermore, continuous phase
distributions are directly recovered so that associated difficulties in the
2D phase unwrapping can be efficiently bypassed. The above-mentioned
advantages of TIE along with the rapid developments in related areas
provide us substantial grounds for believing that TIE now has become a
competitive alternative to its conventional interferometric counterparts
and occupied an important place in the fields of optical imaging and
metrology (a brief comparison between the digital holography and TIE
is given in Table 8 [449]).
Though TIE has shown its great potential in the field of optical imag-
ing and microscopy, there are still several important theoretical and
technical issues deserve to be further studied in the future:
1) The applicability of TIE in the presence of zero intensity and or phase
vortex. This has long been a controversial issue since the inception
of TIE. In 2001, Allen et al. [167] proposed to address this prob-
lem by iterative phase retrieval, which requires to record an addi-
tional intensity image at a large defocus distance. In 2013, Lubk et al.
[450] pointed out the mistake in Allen et al.’s derivation and sug-
gested a new “exhaustive testing” method to rule out the intensity
ambiguities induced by phase vortex. However, in practical experi-
ments, it is very difficult to accurately localize all the singular points
(zero intensities) inside the domain due to the noise and other dis-
turbances.
2) The “deterministicity” of the phase (RI) retrieval depends on certain
approximations. Deterministic phase retrieval relies on the lineariza-
tion of the relationship between the intensity and phase information.
However, the establishment of linearization generally depends on
certain approximations. For example, the traditional TIE is based
on paraxial approximation and the weak defocusing approximation.
The CTF or mixed transfer function methods require the sample to
satisfy the weak object approximation or slowly varying object ap-
proximation. 3D diffraction tomography based on TIE is only valid
for weak scattering samples under first-order Born approximation,
etc.. However, in practical applications, these approximations are
generally not easy to be satisfied strictly.
3) The quantitativeness of phase (RI) recovery remains debatable. Al-
though TIE has proven to be a quantitative phase recovery method,
according to the reported results among the present literature, the
phase measurement accuracy is still not very high, especially when
compared with the interferometric methods. In fact, although TIE
greatly relaxes the stringent requirements on the light source and
image system of interferometry, it imposes stricter requirements for
precise calibrations of axial defocus distance, accurate acquisition of
C. Zuo, J. Li and J. Sun et al.
boundary conditions, and correct implementation of numerical so-
lutions. These factors become a bottleneck of TIE for high-precision
phase measurement in practical engineering applications.
In addition to the above three challenging issues, the following re-
search directions, we would like to highlight the following future re-
search directions.
1) Super-resolved QPI based on TIE. At present, the spatial resolu-
tion of the phase reconstructed by TIE is still limited to the diffrac-
tion limit of the imaging system (coherent diffraction limit under
coherent illuminations, and incoherent diffraction limit under the
matched annular illumination [209]). To the best of our knowledge,
super-resolved (beyond the incoherent diffraction limit) QPI based
on TIE has not been reported so far, which is an important but chal-
lenging direction for future research.
2) 4D light-field retrieval based on “transport of intensity”. For
general partially coherent fields, it is difficult to retrieve its com-
plete reconstruction by TIE due to its inherent high dimensional-
ity (4D). But if we can further incorporate some additional infor-
mation, such as combining the low-resolution field camera (Shack-
Hartmann) measurement with the high-resolution through-focus in-
tensity stack, the amount of information should be sufficient for com-
plete 4D light-field retrieval, or equivalently, WDF/AF retrieval at
full image resolution.
3) Generalized phase retrieval beyond “transport of intensity”. In
TIE, the phase contrast is produced by defocusing (free-space propa-
gation), thereby converting the invisible phase information into vis-
ible intensity signal. Nevertheless, the way to create the phase con-
trast is not just limited to defocusing. Many aberration functions,
such as astigmatism, spherical aberration, and Zernike phase annu-
lus, can create phase contrast even at the in-focus plane, which may
provide more possibilities for designing completely new determinis-
tic phase retrieval methods beyond the “Transport of intensity”.
4) “Deep learning” enabled TIE phase retrieval. The paraxial ap-
proximation, weak defocusing approximation, boundary artifacts,
phase discrepancy, low-frequency noise, and high-frequency phase
blurring are typical limiting factors of TIE, which would probably
be addressed by the recent deep learning technologies [451–455].
Different from the traditional TIE, which first “establishes the for-
ward mathematical model” and then “solves the corresponding in-
verse problem”, depth learning can directly establish the complex
nonlinear relationship between the input (defocused intensity im-
ages) and the ideal output (quantitative phase distribution) with the
help of a large amount of training data, thus opening new avenues
for circumventing the limitations of the conventional TIE.
The four research directions mentioned above are mainly at the the-
oretical level. Of course, developing TIE microscopic imaging systems
towards more integrated, more miniaturized, more easy-to-use, higher
performance, and higher reliability is also an important direction for
future research.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influ-
ence the work reported in this paper. Redistribution and use of MAT-
LAB source code must retain the copyright notice, list of conditions, the
disclaimer from the authors.
CRediT authorship contribution statement
Chao Zuo: Conceptualization, Writing - original draft, Software,
Data curation, Visualization, Supervision, Project administration, Fund-
ing acquisition. Jiaji Li: Investigation, Writing - review & editing.
Jiasong Sun: Investigation, Writing - review & editing. Yao Fan: Data
Optics and Lasers in Engineering 135 (2020) 106187
curation, Writing - review & editing. Jialin Zhang: Software, Writing
- review & editing. Linpeng Lu: Writing - review & editing. Runnan
Zhang: Writing - review & editing. Bowen Wang: Visualization. Lei
Huang: Writing - review & editing. Qian Chen: Resources, Supervision,
Project administration, Funding acquisition.
Acknowledgments
The authors thank J. Ding for help with the schematic drawing.
This work was supported by the National Natural Science Foundation of
China (61722506), Leading Technology of Jiangsu Basic Research Plan
(BK20192003), Final Assembly “13th Five-Year Plan” Advanced Re-
search Project of China (30102070102), Equipment Advanced Research
Fund of China (61404150202), National Defense Science and Technol-
ogy Foundation of China (2019-JCJQ-JJ-381), Outstanding Youth Foun-
dation of Jiangsu Province (BK20 170034), The Key Research and De-
velopment Program of Jiangsu Province (BE2017162), “333 Engineer-
ing” Research Project of Jiangsu Province (BRA2016407), Fundamental
Research Funds for the Central Universities (30920032101), and Open
Research Fund of Jiangsu Key Laboratory of Spectral Imaging & Intelli-
gent Sense (30918 01410411).
Supplementary material
Supplementary material associated with this article can be found, in
the online version, at 10.1016/j.optlaseng.2020.106187
References
[1] Cowley JM. Diffraction physics. Elsevier; 1995.
[2] Goodman JW. Introduction to Fourier Optics. Roberts and Company Publishers;
2005.
[3] Griffiths D.J.. Introduction to electrodynamics. 2005.
[4] Hecht E. Optics. 4 edition. Reading, Mass: Addison-Wesley; 2001. ISBN
978-0-8053-8566-3.
[5] Born M, Wolf E, Bhatia AB, Clemmow PC, Gabor D, Stokes AR, Taylor AM, Way-
man PA, Wilcock WL. Principles of Optics: Electromagnetic Theory of Propagation,
Interference and Diffraction of Light. 7th edition. Cambridge ; New York: Cam-
bridge University Press; 1999. ISBN 978-0-521-64222-4.
[6] High-speed camera. 2017. Page Version ID: 786672531;
[7] Oppenheim AV, Lim JS. The importance of phase in signals. Proceedings of the
IEEE 1981;69(5):529–41.
[8] Giloh H, Sedat JW. Fluorescence microscopy: reduced photobleaching of
rhodamine and fluorescein protein conjugates by n-propyl gallate. Science
1982;217(4566):1252–5.
[9] Webb RH. Confocal optical microscopy. Rep Prog Phys 1996;59(3):427.
[10] Diaspro A, et al. Confocal and two-photon microscopy: foundations, applications,
and advances, vol 1. Wiley-Liss New York; 2002.
[11] Gustafsson MG. Surpassing the lateral resolution limit by a factor of two using
structured illumination microscopy. J Microsc 2000;198(Pt 2):82–7.
[12] Hell SW, Wichmann J. Breaking the diffraction resolution limit by stimu-
lated emission: stimulated-emission-depletion fluorescence microscopy. Opt Lett
1994;19(11):780–2.
[13] Betzig E, Patterson GH, Sougrat R, Lindwasser OW, Olenych S, Bonifacino JS,
Davidson MW, Lippincott-Schwartz J, Hess HF. Imaging intracellular fluorescent
proteins at nanometer resolution. Science 2006;313(5793):1642–5.
[14] Rust MJ, Bates M, Zhuang X. Sub-diffraction-limit imaging by stochas-
tic optical reconstruction microscopy (STORM). Nat Methods 2006;3:793–6.
doi:10.1038/nmeth929.
[15] Stephens DJ, Allan VJ. Light microscopy techniques for live cell imaging. Science
2003;300(5616):82–6.
[16] Zernike F. Phase contrast, a new method for the microscopic observation of trans-
parent objects. Physica 1942;9:686–98.
[17] Nomarski G. Differential microinterferometer with polarized waves; 1955. 9S–11S.
[18] Popescu G. Quantitative Phase Imaging of Cells and Tissues. McGraw-Hill Profes-
sional; 2011.
[19] Park Y, Depeursinge C, Popescu G. Quantitative phase imaging in biomedicine. Nat
Photonics 2018;12(10):578–89.
[20] Abramovici A, Althouse WE, Drever RW, Grsel Y, Kawamura S, Raab FJ, Shoe-
maker D, Sievers L, Spero RE, Thorne KS, Vogt RE, Weiss R, Whitcomb SE,
Zucker ME. Ligo: The laser interferometer gravitational-wave observatory. Science
1992;256(5055):325–33.
[21] Collaboration LS, Collaboration V. Observation of Gravitational Waves from a Bi-
nary Black Hole Merger. Phys Rev Lett 2016;116(6):061102.
[22] Lekberg O. Electronic speckle pattern interferometry. Phys in Tech 1980;11(1):16.
[23] Wang W-C, Hwang C-H, Lin S-Y. Vibration measurement by the time-averaged elec-
tronic speckle pattern interferometry methods. Appl Optics 1996;35(22):4502–9.
C. Zuo, J. Li and J. Sun et al.
[24] Schwarz CJ, Kuznetsova Y, Brueck SRJ. Imaging interferometric microscopy. Opt
Lett 2003;28(16):1424–6.
[25] Popescu G, Ikeda T, Dasari RR, Feld MS. Diffraction phase microscopy for quanti-
fying cell structure and dynamics. Opt Lett 2006;31(6):775–7.
[26] Kuznetsova Y, Neumann A, Brueck SRJ. Imaging interferometric microscopy-
–approaching the linear systems limits of optical resolution. Opt Express
2007;15(11):6651–63.
[27] Cuche E, Bevilacqua F, Depeursinge C. Digital holography for quantitative phase–
contrast imaging. Opt Lett 1999a;24(5):291–3.
[28] Cuche E, Marquet P, Depeursinge C. Simultaneous amplitude-contrast and quan-
titative phase-contrast microscopy by numerical reconstruction of Fresnel off-axis
holograms. Appl Optics 1999b;38(34):6994–7001.
[29] Cuche E, Marquet P, Depeursinge C. Spatial filtering for zero-order and twin-image
elimination in digital off-axis holography. Appl Optics 2000;39(23):4070–5.
[30] Schnars U, Jptner WPO. Digital recording and numerical reconstruction of holo-
grams. Meas Sci Technol 2002;13(9):R85.
[31] Schnars U, Jueptner W. Digital holography: digital hologram recording, numerical
reconstruction, and related techniques. Springer Science & Business Media; 2005.
[32] Marquet P, Rappaz B, Magistretti PJ, Cuche E, Emery Y, Colomb T, Depeursinge C.
Digital holographic microscopy: a noninvasive contrast imaging technique allowing
quantitative visualization of living cells with subwavelength axial accuracy. Opt
Lett 2005;30(5):468–70.
[33] Kemper B, Langehanenberg P, Von Bally G. Digital holographic microscopy: A new
method for surface analysis and marker-free dynamic life cell imaging. Optik Pho-
tonik 2007;2(2):41–4.
[34] Kemper B, Bally Gv. Digital holographic microscopy for live cell applications and
technical inspection. Appl Optics 2008;47(4):A52–61.
[35] Kim MK. Digital holographic microscopy. In: Digital Holographic Microscopy.
Springer; 2011. p. 149–90.
[36] Wang Z, Millet L, Mir M, Ding H, Unarunotai S, Rogers J, Gillette MU, Popescu G.
Spatial light interference microscopy (SLIM). Opt Express 2011;19(2):1016–26.
[37] Bhaduri B, Pham H, Mir M, Popescu G. Diffraction phase microscopy with white
light. Opt Lett 2012;37(6):1094–6.
[38] Bon P, Maucort G, Wattellier B, Monneret S. Quadriwave lateral shearing in-
terferometry for quantitative phase microscopy of living cells. Opt Express
2009;17(15):13080–94.
[39] Girshovitz P, Shaked NT. Compact and portable low-coherence interferometer with
off-axis geometry for quantitative phase microscopy and nanoscopy. Opt Express
2013;21(5):5701–14.
[40] Shack RV. Production and use of a lecticular hartmann screen. J Opt Soc Am
1971;61:656–61.
[41] Hartmann J. Bemerkungen uber den bau und die justirung von spektrographen. Zt
Instrumentenkd 1990;20(47):17–27.
[42] Platt BC, Shack R. History and principles of shack-hartmann wavefront sensing. J
Refract Surg 2001;17(5):S573–7.
[43] Ragazzoni R. Pupil plane wavefront sensing with an oscillating prism. J Mod Opt
1996;43(2):289–93.
[44] Esposito S, Riccardi A. Pyramid wavefront sensor behavior in partial correction
adaptive optic systems. Astron Astrophys 2001;369(2):L9–L12.
[45] Ragazzoni R, Diolaiti E, Vernet E. A pyramid wavefront sensor with no dynamic
modulation. Opt Commun 2002;208(1-3):51–60.
[46] Neil MAA, Booth MJ, Wilson T. New modal wave-front sensor: a theoretical anal-
ysis. J Opt Soc Am A 2000;17:1098–107.
[47] Booth MJ. Wave front sensor-less adaptive optics: a model-based approach using
sphere packings. Opt Express 2006;14:1339–52.
[48] Schfer B, Mann K. Determination of beam parameters and coherence properties
of laser radiation by use of an extended hartmann-shack wave-front sensor. Appl
Optics 2002;41:2809–17.
[49] Schfer B, Lbbecke M, Mann K. Hartmann-shack wave front measurements for
real time determination of laser beam propagation parameters. Rev Sci Instrum
2006;77:053103.
[50] Pfund J, Lindlein N, Schwider J, Burow R, Blmel T, Elssner KE. Absolute sphericity
measurement: a comparative study of the use of interferometry and a shack-hart-
mann sensor. Opt Lett 1998;23:742–4.
[51] Greivenkamp JE, Smith DG, Gappinger RO, Williby GA. Optical testing using shack-
-hartmann wavefront sensors. In: Optical Engineering for Sensing and Nanotech-
nology (ICOSN 2001), vol. 4416. International Society for Optics and Photonics;
2001. p. 260–3.
[52] Dayton D, Gonglewski J, Pierson B, Spielbusch B. Atmospheric structure
function measurements with a Shack-Hartmann wave-front sensor. Opt Lett
1992;17:1737–9.
[53] Ricklin JC, Davidson FM. Atmospheric turbulence effects on a partially coherent
Gaussian beam: implications for free-space laser communication. J Opt Soc Am A
2002;19:1794–802.
[54] Booth MJ. Adaptive optics in microscopy. Philos Trans R Soc Lond Math Phys Eng
Sci 2007;365:2829–43.
[55] Cha JW, Ballesta J, So PTC. Shack-hartmann wavefront-sensor-based
adaptive optics system for multiphoton microscopy. J Biomed Opt
2010;15:046022–046022–10.
[56] Liang J, Grimm B, Goelz S, Bille JF. Objective measurement of wave aberrations of
the human eye with the use of a hartmann-shack wave-front sensor. J Opt Soc Am
A 1994;11:1949–57.
[57] Moreno-Barriuso E, Navarro R. Laser ray tracing versus hartmann-shack sensor for
measuring optical aberrations in the human eye. J Opt Soc Am A 2000;17:974–85.
[58] Kohnen T, Koch DD, et al. Cataract and refractive surgery. Springer; 2006.
Optics and Lasers in Engineering 135 (2020) 106187
[59] Gerchberg RW. Phase determination from image and diffraction plane pictures in
the electron microscope. Optik 1971;34:275–84.
[60] Gerchberg RW, Saxton WO. A practical algorithm for the determination of the phase
from image and diffraction plane pictures. Optik (Jena) 1972;35:237.
[61] Miao J, Charalambous P, Kirz J, Sayre D. Extending the methodology of X-ray
crystallography to allow imaging of micrometre-sized non-crystalline specimens.
Nature 1999;400(6742):342–4.
[62] Zuo JM, Vartanyants I, Gao M, Zhang R, Nagahara LA. Atomic resolution imaging of
a carbon nanotube from diffraction intensities. Science 2003;300(5624):1419–21.
[63] Marchesini S, He H, Chapman HN, Hau-Riege SP, Noy A, Howells MR, Weierstall U,
Spence JCH. X-ray image reconstruction from a diffraction pattern alone. Phys Rev
B 2003;68:140101.
[64] Eisebitt S, Lning J, Schlotter WF, Lrgen M, Hellwig O, Eberhardt W, Sthr J.
Lensless imaging of magnetic nanostructures by X-ray spectro-holography. Nature
2004;432:885–8.
[65] Gonsalves RA, Chidlaw R. Wavefront sensing by phase retrieval. In: Applications
of Digital Image Processing III, vol. 207. International Society for Optics and Pho-
tonics; 1979. p. 32–9.
[66] Guyon O. Limits of adaptive optics for high-contrast imaging. ApJ
2005;629(1):592.
[67] Pedrini G, Osten W, Zhang Y. Wave-front reconstruction from a sequence of inter-
ferograms recorded at different planes. Opt Lett 2005;30(8):833–5.
[68] Anand A, Pedrini G, Osten W, Almoro P. Wavefront sensing with random amplitude
mask and phase retrieval. Opt Lett 2007;32(11):1584–6.
[69] Almoro PF, Pedrini G, Gundu PN, Osten W, Hanson SG. Phase microscopy of tech-
nical and biological samples through random phase modulation with a diffuser.
Opt Lett 2010;35(7):1028–30.
[70] Mudanyali O, Tseng D, Oh C, Serhan OI, Sencan I, Bishara W, Oztoprak C, Seo S,
Khademhosseini B, Ozcan A. Compact, light-weight and cost-effective microscope
based on lensless incoherent holography for telemedicine applications. Lab Chip
2010;10(11):1417–28.
[71] Tseng D, Mudanyali O, Oztoprak C, Serhan OI, Sencan I, Yaglidere O, Ozcan A.
Lensfree microscopy on a cellphone. Lab Chip 2010;10(14):1787–92.
[72] Wackerman CC, Yagle AE. Use of fourier domain real-plane zeros to overcome a
phase retrieval stagnation. J Opt Soc Am A 1991;8(12):1898–904.
[73] Lu G, Zhang Z, Yu FTS, Tanone A. Pendulum iterative algorithm for phase retrieval
from modulus data. Opt Eng 1994;33(2):548–55.
[74] Takajo H, Takahashi T, Kawanami H, Ueda R. Numerical investigation of the it-
erative phase-retrieval stagnation problem: territories of convergence objects and
holes in their boundaries. J Opt Soc Am A 1997;14(12):3175–87.
[75] Misell DL. A method for the solution of the phase problem in electron microscopy.
J Phys D: Appl Phys 1973;6(1):L6.
[76] Allen LJ, Oxley MP. Phase retrieval from series of images obtained by defocus
variation. Opt Commu 2001;199:65–75.
[77] Zhang Y, Pedrini G, Osten W, Tiziani HJ. Whole optical wave field reconstruction
from double or multi in-line holograms by phase retrieval algorithm. Opt Express
2003;11(24):3234–41.
[78] Bao P, Zhang F, Pedrini G, Osten W. Phase retrieval using multiple illumination
wavelengths. Opt Lett 2008;33(4):309–11.
[79] Zhang F, Pedrini G, Osten W. Phase retrieval of arbitrary complex-valued fields
through aperture-plane modulation. Phys Rev A 2007;75(4):043805.
[80] Fannjiang A, Liao W. Phase retrieval with random phase illumination. J Opt Soc
Am A 2012;29(9):1847–59.
[81] Almoro P, Pedrini G, Osten W. Complete wavefront reconstruction using
sequential intensity measurements of a volume speckle field. Appl Optics
2006;45(34):8596–605.
[82] Gao P, Pedrini G, Zuo C, Osten W. Phase retrieval using spatially modulated illu-
mination. Opt Lett 2014;39(12):3615.
[83] Bates RHT. Uniqueness of solutions to two-dimensional fourier phase problems for
localized and positive images. Computer Vision, Graphics, and Image Processing
1984;25(2):205–17.
[84] Fienup JR, Wackerman CC. Phase-retrieval stagnation problems and solutions. J
Opt Soc Am A 1986;3(11):1897–907.
[85] Seldin JH, Fienup JR. Numerical investigation of the uniqueness of phase retrieval.
J Opt Soc Am A 1990;7(3):412–27.
[86] Guizar-Sicairos M, Fienup JR. Understanding the twin-image problem in phase re-
trieval. J Opt Soc Am A 2012;29(11):2367–75.
[87] Fienup JR. Phase retrieval algorithms: a comparison. Appl Optics
1982;21(15):2758–69.
[88] Fienup JR. Reconstruction of an object from the modulus of its Fourier transform.
Opt Lett 1978;3(1):27–9.
[89] Bauschke HH, Combettes PL, Luke DR. Hybrid projection-reflection method for
phase retrieval. J Opt Soc Am A 2003;20(6):1025–34.
[90] Elser V. Phase retrieval by iterated projections. J Opt Soc Am A 2003;20(1):40–55.
[91] Luke DR. Relaxed averaged alternating reflections for diffraction imaging. Inverse
Problems 2005;21(1):37–50.
[92] Oszlányi G, Sütő A. Ab initio structure solution by charge flipping. Acta Crystallogr
A 2004;60(2):134–41.
[93] Marchesini S. Invited article: A unified evaluation of iterative projection algorithms
for phase retrieval. Rev Sci Instrum 2007;78(1):011301.
[94] Hegerl R, Hoppe W. Dynamische Theorie der Kristallstrukturanalyse durch Elek-
tronenbeugung im inhomogenen Primrstrahlwellenfeld. Berichte der Bunsenge-
sellschaft fr physikalische Chemie 1970;74(11):1148–54.
[95] Faulkner HML, Rodenburg JM. Movable aperture lensless transmission microscopy:
A novel phase retrieval algorithm. Phys Rev Lett 2004;93(2):023903.
C. Zuo, J. Li and J. Sun et al.
[96] Faulkner HML, Rodenburg JM. Error tolerance of an iterative phase re-
trieval algorithm for moveable illumination microscopy. Ultramicroscopy
2005;103(2):153–64.
[97] Guizar-Sicairos M, Fienup JR. Phase retrieval with transverse translation diversity:
a nonlinear optimization approach. Opt Express 2008;16(10):7264–78.
[98] Thibault P, Dierolf M, Menzel A, Bunk O, David C, Pfeiffer F. High-resolution scan-
ning x-ray diffraction microscopy. Science 2008;321(5887):379–82.
[99] Maiden AM, Rodenburg JM. An improved ptychographical phase retrieval algo-
rithm for diffractive imaging. Ultramicroscopy 2009;109(10):1256–62.
[100] Thibault P, Dierolf M, Bunk O, Menzel A, Pfeiffer F. Probe retrieval in ptycho-
graphic coherent diffractive imaging. Ultramicroscopy 2009;109(4):338–43.
[101] Thibault P, Guizar-Sicairos M. Maximum-likelihood refinement for coherent
diffractive imaging. New J Phys 2012;14(6):063004.
[102] Maiden A, Johnson D, Li P. Further improvements to the ptychographical iterative
engine. Optica 2017;4(7):736–45.
[103] Maiden AM, Humphry MJ, Sarahan MC, Kraus B, Rodenburg JM. An anneal-
ing algorithm to correct positioning errors in ptychography. Ultramicroscopy
2012a;120:64–72.
[104] Beckers M, Senkbeil T, Gorniak T, Giewekemeyer K, Salditt T, Rosenhahn A. Drift
correction in ptychographic diffractive imaging. Ultramicroscopy 2013;126:44–7.
[105] Zhang F, Peterson I, Vila-Comamala J, Diaz A, Berenguer F, Bean R, Chen B, Men-
zel A, Robinson IK, Rodenburg JM. Translation position determination in ptycho-
graphic coherent diffraction imaging. Opt Express 2013;21(11):13592.
[106] Thibault P, Menzel A. Reconstructing state mixtures from diffraction measure-
ments. Nature 2013;494(7435):68–71.
[107] Batey DJ, Claus D, Rodenburg JM. Information multiplexing in ptychography. Ul-
tramicroscopy 2014;138:13–21.
[108] Clark JN, Huang X, Harder RJ, Robinson IK. Dynamic imaging using ptychography.
Phys Rev Lett 2014;112(11):113901.
[109] Karl R, Bevis C, Lopez-Rios R, Reichanadter J, Gardner D, Porter C, Shanblatt E,
Tanksalvala M, Mancini GF, Murnane M, Kapteyn H, Adams D. Spatial, spectral,
and polarization multiplexed ptychography. Opt Express 2015;23(23):30250.
[110] Maiden AM, Humphry MJ, Zhang F, Rodenburg JM. Superresolution imaging via
ptychography. J Opt Soc Am A 2011;28(4):604–12.
[111] Humphry MJ, Kraus B, Hurst AC, Maiden AM, Rodenburg JM. Ptychographic elec-
tron microscopy using high-angle dark-field scattering for sub-nanometre resolu-
tion imaging. Nat Commun 2012;3:730.
[112] Stockmar M, Cloetens P, Zanette I, Enders B, Dierolf M, Pfeiffer F, Thibault P.
Near-field ptychography: phase retrieval for inline holography using a structured
illumination. Sci Rep 2013;3(1).
[113] Takahashi Y, Suzuki A, Furutaku S, Yamauchi K, Kohmura Y, Ishikawa T. High-
-resolution and high-sensitivity phase-contrast imaging by focused hard x-ray pty-
chography with a spatial filter. Appl Phys Lett 2013;102(9):094102.
[114] Maiden AM, Humphry MJ, Rodenburg JM. Ptychographic transmission mi-
croscopy in three dimensions using a multi-slice approach. J Opt Soc Am A
2012b;29(8):1606–14.
[115] Godden TM, Suman R, Humphry MJ, Rodenburg JM, Maiden AM. Ptychographic
microscope for three-dimensional imaging. Opt Express 2014;22(10):12513.
[116] Suzuki A, Furutaku S, Shimomura K, Yamauchi K, Kohmura Y, Ishikawa T, Taka-
hashi Y. High-resolution multislice x-ray ptychography of extended thick objects.
Phys Rev Lett 2014;112(5).
[117] Shimomura K, Suzuki A, Hirose M, Takahashi Y. Precession x-ray ptychography
with multislice approach. Phys Rev B 2015;91(21).
[118] Rodenburg JM, Hurst AC, Cullis AG, Dobson BR, Pfeiffer F, Bunk O, David C, Jefi-
movs K, Johnson I. Hard-x-ray lensless imaging of extended objects. Phys Rev Lett
2007a;98(3):034801.
[119] Giewekemeyer K, Thibault P, Kalbfleisch S, Beerlink A, Kewish CM, Dierolf M,
Pfeiffer F, Salditt T. Quantitative biological imaging by ptychographic x-ray
diffraction microscopy. Proceedings of the National Academy of Sciences
2010;107(2):529–34.
[120] Maiden AM, Morrison GR, Kaulich B, Gianoncelli A, Rodenburg JM. Soft X-ray
spectromicroscopy using ptychography with randomly phased illumination. Nat
Commun 2013;4:1669.
[121] Rodenburg JM, Hurst AC, Cullis AG. Transmission microscopy without lenses for
objects of unlimited size. Ultramicroscopy 2007b;107(2-3):227–31.
[122] Hue F, Rodenburg JM, Maiden AM, Sweeney F, Midgley PA. Wave-front phase
retrieval in transmission electron microscopy via ptychography. Phys Rev B
2010;82(12).
[123] Hue F, Rodenburg JM, Maiden AM, Midgley PA. Extended ptychography in the
transmission electron microscope: Possibilities and limitations. Ultramicroscopy
2011;111(8):1117–23.
[124] Brady GR, Guizar-Sicairos M, Fienup JR. Optical wavefront measurement
using phase retrieval with transverse translation diversity. Opt Express
2009;17(2):624–39.
[125] Maiden AM, Rodenburg JM, Humphry MJ. Optical ptychography: a practical im-
plementation with useful resolution. Opt Lett 2010;35(15):2585–7.
[126] Marrison J, Rty L, Marriott P, O’Toole P. Ptychography - a label free, high-con-
trast imaging technique for live cells using quantitative phase information. Sci Rep
2013;3(1).
[127] Zheng G, Horstmeyer R, Yang C. Wide-field, high-resolution Fourier ptychographic
microscopy. Nat Photonics 2013;7(9):739–45.
[128] Ou X, Zheng G, Yang C. Embedded pupil function recovery for Fourier ptycho-
graphic microscopy. Opt Express 2014;22(5):4960.
[129] Sun J, Chen Q, Zhang Y, Zuo C. Efficient positional misalignment correction method
for Fourier ptychographic microscopy. Biomed Opt Express 2016a;7(4):1336.
Optics and Lasers in Engineering 135 (2020) 106187
[130] Yeh L-H, Dong J, Zhong J, Tian L, Chen M, Tang G, Soltanolkotabi M, Waller L.
Experimental robustness of Fourier ptychography phase retrieval algorithms. Opt
Express 2015;23(26):33214.
[131] Dong S, Shiradkar R, Nanda P, Zheng G. Spectral multiplexing and coheren-
t-state decomposition in fourier ptychographic imaging. Biomed Opt Express
2014;5(6):1757.
[132] Tian L, Li X, Ramchandran K, Waller L. Multiplexed coded illumination for
fourier ptychography with an LED array microscope. Biomed Opt Express
2014;5(7):2376–89.
[133] Sun J, Chen Q, Zhang Y, Zuo C. Sampling criteria for fourier ptychographic mi-
croscopy in object space and frequency space. Opt Express 2016b;24(14):15765.
[134] Li P, Batey DJ, Edo TB, Rodenburg JM. Separation of three-dimensional scattering
effects in tilt-series fourier ptychography. Ultramicroscopy 2015;158:1–7.
[135] Tian L, Waller L. 3d intensity and phase imaging from light field measurements in
an LED array microscope. Optica 2015a;2(2):104.
[136] Horstmeyer R, Chen RY, Ou X, Ames B, Tropp JA, Yang C. Solving ptychography
with a convex relaxation. New J Phys 2015;17(5):053044.
[137] Zuo C, Sun J, Chen Q. Adaptive step-size strategy for noise-robust Fourier ptycho-
graphic microscopy. Opt Express 2016a;24(18):20724.
[138] Sun J, Zuo C, Zhang L, Chen Q. Resolution-enhanced Fourier ptychographic mi-
croscopy based on high-numerical-aperture illuminations. Sci Rep 2017;7(1):1187.
[139] Williams GJ, Quiney HM, Peele AG, Nugent KA. Coherent diffractive imag-
ing and partial coherence. Phys Rev B 2007;75:104102. doi:10.1103/Phys-
RevB.75.104102.
[140] Teague MR. Irradiance moments: their propagation and use for unique retrieval of
phase. J Opt Soc Am 1982;72(9):1199–209.
[141] Teague MR. Deterministic phase retrieval: a green’s function solution. J Opt Soc
Am A 1983;73(11):1434–41.
[142] Streibl N. Phase imaging by the transport equation of intensity. Opt Commun
1984;49(1):6–10.
[143] Ichikawa K, Lohmann AW, Takeda M. Phase retrieval based on the irradiance
transport equation and the Fourier transform method: experiments. Appl Optics
1988;27(16):3433–6.
[144] Roddier F. Wavefront sensing and the irradiance transport equation. Appl Optics
1990;29(10):1402–3.
[145] Roddier F, Roddier C, Roddier N. Curvature Sensing: A New Wavefront Sensing
Method. In: Morris GM, editor. Statistical Optics, vol. 0976. SPIE; 1988. p. 203–9.
International Society for Optics and Photonics
[146] Roddier F. Curvature sensing and compensation: a new concept in adaptive optics.
Appl Optics 1988;27(7):1223–5.
[147] Roddier NA. Algorithms for wavefront reconstruction out of curvature sensing data.
In: Active and Adaptive Optical Systems, vol. 1542. International Society for Optics
and Photonics; 1991. p. 120–9.
[148] Roddier C, Roddier F. Wave-front reconstruction from defocused images and the
testing of ground-based optical telescopes. J Opt Soc Am A 1993;10(11):2277–87.
[149] Roddier F, Roddier C. Wavefront reconstruction using iterative fourier transforms.
Appl Optics 1991;30:1325–7.
[150] Roddier F, Northcott M, Graves JE. A simple low-order adaptive optics system for
near-infrared applications. Publ Astron Soc Pac 1991;103(659):131.
[151] Roddier FJ, Anuskiewicz J, Graves JE, Northcott MJ, Roddier CA. Adaptive optics
at the university of hawaii i: current performance at the telescope. In: Adaptive
Optics in Astronomy, vol. 2201. International Society for Optics and Photonics;
1994. p. 2–9.
[152] Graves JE, Northcott MJ, Roddier FJ, Roddier CA, Close LM. First light for
hokupa’a: 36-element curvature AO system at UH. In: Adaptive Optical System
Technologies, vol. 3353. International Society for Optics and Photonics; 1998.
p. 34–43.
[153] Arsenault R, Alonso J, Bonnet H, Brynnel J, Delabre B, Donaldson R, Dupuy C,
Fedrigo E, Farinato J, Hubin NN, et al. Macao-vlti: an adaptive optics system for
the eso interferometer. In: Adaptive Optical System Technologies II, vol. 4839.
International Society for Optics and Photonics; 2003. p. 174–85.
[154] Hayano Y, Takami H, Guyon O, Oya S, Hattori M, Saito Y, Watanabe M, Mu-
rakami N, Minowa Y, Ito M, et al. Current status of the laser guide star adaptive
optic system for subaru telescope. In: Adaptive Optics Systems, vol. 7015. Interna-
tional Society for Optics and Photonics; 2008. p. 701510.
[155] Gureyev TE, Roberts A, Nugent KA. Partially coherent fields, the transport-of-inten-
sity equation, and phase uniqueness. J Opt Soc Am A, JOSA A 1995a;12(9):1942–6.
[156] Gureyev TE, Roberts A, Nugent KA. Phase retrieval with the transport-of-inten-
sity equation: matrix solution with use of Zernike polynomials. J Opt Soc Am A
1995b;12(9):1932–41.
[157] Gureyev TE, Nugent KA. Phase retrieval with the transport-of-intensity equa-
tion. II. orthogonal series solution for nonuniform illumination. J Opt Soc Am A
1996;13(8):1670–82.
[158] Gureyev TE, Nugent KA. Rapid quantitative phase imaging using the transport of
intensity equation. Opt Commun 1997;133(1):339–46.
[159] Paganin D, Nugent KA. Noninterferometric phase imaging with partially coherent
light. Phys Rev Lett 1998;80(12):2586.
[160] Snigirev A, Snigireva I, Kohn V, Kuznetsov S, Schelokov I. On the possibilities of
x-ray phase contrast microimaging by coherent high-energy synchrotron radiation.
Rev Sci Instrum 1995;66(12):5486–92.
[161] Nugent KA, Gureyev TE, Cookson DF, Paganin D, Barnea Z. Quantitative phase
imaging using hard x rays. Phy Rev Lett 1996;77:2961.
[162] Wilkins SW, Gureyev TE, Gao D, Pogany A, Stevenson AW. Phase-contrast imaging
using polychromatic hard x-rays. Nature 1996;384(6607):335.
C. Zuo, J. Li and J. Sun et al.
[163] Cloetens P, Ludwig W, Baruchel J, Van Dyck D, Van Landuyt J, Guigay JP,
Schlenker M. Holotomography: Quantitative phase tomography with microm-
eter resolution using hard synchrotron radiation x rays. Appl Phys Lett
1999;75(19):2912–14.
[164] Allman BE, McMahon PJ, Nugent KA, Paganin D, Jacobson DL, Arif M, Werner SA.
Phase radiography with neutrons. Nature 2000;408:158.
[165] McMahon PJ, Allman BE, Jacobson DL, Arif M, Werner SA, Nugent KA.
Quantitative phase radiography with polychromatic neutrons. Phys Rev Lett
2003;91(14):145502.
[166] Bajt S, Barty A, Nugent KA, McCartney M, Wall M, Paganin D. Quantitative
phase-sensitive imaging in a transmission electron microscope. Ultramicroscopy
2000;83(1-2):67–73.
[167] Allen LJ, Faulkner HML, Nugent KA, Oxley MP, Paganin D. Phase retrieval from
images in the presence of first-order vortices. Phys Rev E 2001a;63(3):037602.
[168] Allen LJ, Faulkner HML, Oxley MP, Paganin D. Phase retrieval and aberration
correction in the presence of vortices in high-resolution transmission electron mi-
croscopy. Ultramicroscopy 2001b;88(2):85–97.
[169] McMahon PJ, Barone-Nugent ED, Allman BE, Nugent KA. Quantitative phase-am-
plitude microscopy II: differential interference contrast imaging for biological TEM.
J Microsc 2002;206(3):204–8.
[170] Volkov VV, Zhu Y. Phase imaging and nanoscale currents in phase objects imaged
with fast electrons. Phys Rev Lett 2003;91(4):043904.
[171] Beleggia M, Schofield MA, Volkov VV, Zhu Y. On the transport of intensity tech-
nique for phase retrieval. Ultramicroscopy 2004;102(1):37–49.
[172] Volkov VV, Zhu Y. Lorentz phase microscopy of magnetic materials. Ultrami-
croscopy 2004;98(2):271–81.
[173] McVitie S, Cushley M. Quantitative fresnel lorentz microscopy and the transport of
intensity equation. Ultramicroscopy 2006;106(4-5):423–31.
[174] Petersen TC, Keast VJ, Paganin DM. Quantitative TEM-based phase retrieval
of mgo nano-cubes using the transport of intensity equation. Ultramicroscopy
2008;108(9):805–15.
[175] Barty A, Nugent KA, Paganin D, Roberts A. Quantitative optical phase microscopy.
Opt Lett 1998;23(11):817–19.
[176] Barty A, Nugent KA, Roberts A, Paganin D. Quantitative phase tomography. Opt
Commun 2000;175(4-6):329–36.
[177] Barone-Nugent ED, Barty A, Nugent KA. Quantitative phase-amplitude microscopy
I: optical microscopy. J Microsc 2002;206(3):194–203.
[178] Streibl N. Three-dimensional imaging by a microscope. J Opt Soc Am A, JOSAA
1985;2(2):121–7.
[179] Sheppard CJR. Three-dimensional phase imaging with the intensity transport equa-
tion. Appl Optics 2002;41:5951–5.
[180] Curl CL, Bellair CJ, Harris PJ, Allman BE, Roberts A, Nugent KA, Delbridge LMD.
Quantitative phase microscopy: a new tool for investigating the structure and func-
tion of unstained live cells. Clin Exp Pharmacol Physiol 2004;31(12):896–901.
[181] Curl CL, Bellair CJ, Harris PJ, Allman BE, Roberts A, Nugent KA, Delbridge LMD.
Single cell volume measurement by quantitative phase microscopy (QPM): a case
study of erythrocyte morphology. Cell Physiol Biochem 2006;17(5-6):193–200.
[182] Ross GJ, Bigelow AW, Randers-Pehrson G, Peng CC, Brenner DJ. Phase-based cell
imaging techniques for microbeam irradiations. Nucl Instrum Methods Phys Res
Sect B Beam Interact Mater At 2005;241(1):387–91.
[183] Dragomir NM, Goh XM, Curl CL, Delbridge LMD, Roberts A. Quantitative polarized
phase microscopy for birefringence imaging. Opt Express 2007;15(26):17690–8.
[184] Roberts A, Ampem-Lassen E, Barty A, Nugent KA, Baxter GW, Dragomir NM, Hunt-
ington ST. Refractive-index profiling of optical fibers with axial symmetry by use
of quantitative phase microscopy. Opt Lett 2002;27(23):2061–3.
[185] Ampem-Lassen E, Huntington ST, Dragomir NM, Nugent KA, Roberts A. Refractive
index profiling of axially symmetric optical fibers: a new technique. Opt Express
2005;13(9):3277–82.
[186] Dorrer C, Zuegel JD. Optical testing using the transport-of-intensity equation. Opt
Express 2007;15(12):7165–75.
[187] Darudi A, Shomali R, Tavassoly MT. Determination of the refractive index profile of
a symmetric fiber preform by the transport of intensity equation. Opt Laser Technol
2008;40(6):850–3.
[188] Kou SS, Waller L, Barbastathis G, Sheppard CJR. Transport-of-intensity approach to
differential interference contrast (TI-DIC) microscopy for quantitative phase imag-
ing. Opt Lett 2010;35(3):447–9.
[189] Waller L, Tian L, Barbastathis G. Transport of intensity phase-amplitude imaging
with higher order intensity derivatives. Opt Express 2010a;18(12):12552–61.
[190] Waller L, Luo Y, Yang SY, Barbastathis G. Transport of intensity phase imaging in
a volume holographic microscope. Opt Lett 2010b;35(17):2961–3.
[191] Waller L, Kou SS, Sheppard CJR, Barbastathis G. Phase from chromatic aberrations.
Opt Express 2010c;18(22):22817–25.
[192] Kou SS, Waller L, Barbastathis G, Marquet P, Depeursinge C, Sheppard CJR. Quan-
titative phase restoration by direct inversion using the optical transfer function.
Opt Lett 2011;36(14):2671–3.
[193] Almoro PF, Waller L, Agour M, Falldorf C, Pedrini G, Osten W, Hanson SG. En-
hanced deterministic phase retrieval using a partially developed speckle field. Opt
Lett 2012;37(11):2088–90.
[194] Gorthi SS, Schonbrun E. Phase imaging flow cytometry using a focus-stack collect-
ing microscope. Opt Lett 2012;37:707–9.
[195] Waller L, Tsang M, Ponda S, Yang SY, Barbastathis G. Phase and amplitude imaging
from noisy images by Kalman filtering. Opt Express 2011;19(3):2805–15.
[196] Xue B, Zheng S, Cui L, Bai X, Zhou F. Transport of intensity phase imag-
ing from multiple intensities measured in unequally-spaced planes. Opt Express
2011;19(21):20244–50.
Optics and Lasers in Engineering 135 (2020) 106187
[197] Bie R, Yuan X-H, Zhao M, Zhang L. Method for estimating the axial intensity deriva-
tive in the TIE with higher order intensity derivatives and noise suppression. Opt
Express 2012;20(7):8186–91.
[198] Zheng S, Xue B, Xue W, Bai X, Zhou F. Transport of intensity phase imaging
from multiple noisy intensities measured in unequally-spaced planes. Opt Express
2012;20(2):972–85.
[199] Martinez-Carranza J, Falaggis K, Kozacki T. Optimum measurement criteria for the
axial derivative intensity used in transport of intensity-equation-based solvers. Opt
Lett 2014a;39(2):182–5.
[200] Falaggis K, Kozacki T, Kujawinska M. Optimum plane selection criteria for sin-
gle-beam phase retrieval techniques based on the contrast transfer function. Opt
Lett 2014;39(1):30–3.
[201] Zuo C, Chen Q, Asundi A. Boundary-artifact-free phase retrieval with the trans-
port of intensity equation: fast solution with use of discrete cosine transform. Opt
Express 2014a;22(8):9220.
[202] Zuo C, Chen Q, Li H, Qu W, Asundi A. Boundary-artifact-free phase retrieval with
the transport of intensity equation II: applications to microlens characterization.
Opt Express 2014b;22(15):18310.
[203] Huang L, Zuo C, Idir M, Qu W, Asundi A. Phase retrieval with the transport-of-in-
tensity equation in an arbitrarily shaped aperture by iterative discrete cosine trans-
forms. Opt Lett 2015;40(9):1976.
[204] Zuo C, Chen Q, Huang L, Asundi A. Phase discrepancy analysis and compensation
for fast Fourier transform based solution of the transport of intensity equation. Opt
Express 2014c;22(14):17172.
[205] Zuo C, Chen Q, Yu Y, Asundi A. Transport-of-intensity phase imaging us-
ing Savitzky-Golay differentiation filter - theory and applications. Opt Express
2013a;21(5):5346–62.
[206] Sun J, Zuo C, Chen Q. Iterative optimum frequency combination method for high
efficiency phase imaging of absorptive objects based on phase transfer function.
Opt Express 2015;23(21):28031.
[207] Zuo C, Chen Q, Asundi A. Light field moment imaging: comment. Opt Lett
2014d;39(3):654.
[208] Zuo C, Chen Q, Tian L, Waller L, Asundi A. Transport of intensity phase retrieval and
computational imaging for partially coherent fields: The phase space perspective.
Opt Lasers Eng 2015a;71:20–32.
[209] Zuo C, Sun J, Li J, Zhang J, Asundi A, Chen Q. High-resolution transport-of-intensity
quantitative phase microscopy with annular illumination. Sci Rep 2017;7(1):7654.
[210] Li J, Chen Q, Zhang J, Zhang Y, Lu L, Zuo C. Efficient quantitative phase mi-
croscopy using programmable annular LED illumination. Biomed Opt Express
2017a;8(10):4687–705.
[211] Li J, Chen Q, Sun J, Zhang J, Pan X, Zuo C. Optimal illumination pat-
tern for transport-of-intensity quantitative phase microscopy. Opt Express
2018a;26(21):27599–614.
[212] Zuo C, Sun J, Zhang J, Hu Y, Chen Q. Lensless phase microscopy and diffraction
tomography with multi-angle and multi-wavelength illuminations using a LED ma-
trix. Opt Express 2015b;23(11):14314–28.
[213] Li J, Chen Q, Zhang J, Zhang Z, Zhang Y, Zuo C. Optical diffraction tomography
microscopy with transport of intensity equation using a light-emitting diode array.
Opt Lasers Eng 2017b;95:26–34.
[214] Li J, Chen Q, Sun J, Zhang J, Ding J, Zuo C. Three-dimensional tomographic mi-
croscopy technique with multi-frequency combination with partially coherent illu-
minations. Biomed Opt Express 2018b;9(6):2526–42.
[215] Li J, Chen Q, Sun J, Zhang J, Zuo C. Multimodal computational microscopy based
on transport of intensity equation. J Biomed Opt 2016;21(12):126003.
[216] Zuo C, Chen Q, Qu W, Asundi A. Noninterferometric single-shot quantitative phase
microscopy. Opt Lett 2013b;38(18):3538.
[217] Zuo C, Chen Q, Qu W, Asundi A. High-speed transport-of-intensity phase mi-
croscopy with an electrically tunable lens. Opt Express 2013c;21(20):24060.
[218] Kirchhoff G. Zur theorie der lichtstrahlen. Ann Phys 1883;254(4):663–95.
[219] Rayleigh L. On the passage of waves through apertures in Plane screens, and allied
problems. The London, Edinburgh, and Dublin Philosophical Magazine and Journal
of Science 1897;43(263):259–72. Publisher: Taylor & Francis
[220] Sommerfeld A. Mathematische theorie der diffraction. Math Ann
1896;47(2):317–74.
[221] Guigay JP. Fourier transform analysis of fresnel diffraction patterns and in-line
holograms. Optik (Stuttgart) 1977;49:121–5.
[222] Gilbarg D, Trudinger NS. Elliptic partial differential equations of second order.
springer; 2015.
[223] Nugent KA. Partially coherent diffraction patterns and coherence measurement. J
Opt Soc Am A 1991;8(10):1574–9.
[224] Bhamra KS, et al. Partial differential equations. PHI Learning Pvt Ltd; 2010.
[225] Courant R, Hilbert D. Methods of Mathematical Physics: Partial Differential Equa-
tions. John Wiley & Sons; 2008.
[226] Gureyev TE, Raven C, Snigirev A, Snigireva I, Wilkins SW. Hard x-ray quantitative
non-interferometric phase-contrast microscopy. J Phys Appl Phys 1999;32(5):563.
[227] Pinhasi SV, Alimi R, Perelmutter L, Eliezer S. Topography retrieval using different
solutions of the transport intensity equation. J Opt Soc Am A 2010;27(10):2285–92.
[228] Xue B, Zheng S. Phase retrieval using the transport of intensity equation solved by
the FMG-CG method. Opt - Int J Light Electron Opt 2011;122(23):2101–6.
[229] Voitsekhovich VV. Phase-retrieval problem and orthogonal expansions: curvature
sensing. J Opt Soc Am A 1995;12(10):2194–202.
[230] Ros S, Acosta E, Bar S. Modal phase estimation from wavefront curvature sensing.
Opt Commun 1996;123(4):453–6.
[231] Volkov VV, Zhu Y, De Graef M. A new symmetrized solution for phase retrieval
using the transport of intensity equation. Micron 2002;33(5):411–16.
C. Zuo, J. Li and J. Sun et al.
[232] Frank J, Altmeyer S, Wernicke G. Non-interferometric, non-iterative phase retrieval
by green’s functions. J Opt Soc Am A 2010;27(10):2244–51.
[233] Parvizi A, Müller J, Funken SA, Koch CT. A practical way to resolve ambiguities in
wavefront reconstructions by the transport of intensity equation. Ultramicroscopy
2015;154:1–6.
[234] Martinez-Carranza J, Falaggis K, Kozacki T, Kujawinska M. Effect of imposed
boundary conditions on the accuracy of transport of intensity equation based
solvers. Modeling Aspects in Optical Metrology IV, vol 8789. International Soci-
ety for Optics and Photonics; 2013. 87890N.
[235] Acosta E, Ros S, Soto M, Voitsekhovich VV. Role of boundary measurements in
curvature sensing. Opt Commun 1999;169(1):59–62.
[236] Campbell C. Wave-front sensing by use of a green’s function solution to the intensity
transport equation: comment. J Opt Soc Am A 2007;24(8):2480–1.
[237] Woods SC, Campbell HI, Greenaway AH. Wave-front sensing by use of a Green’s
function solution to the intensity transport equation: reply to comment. J Opt Soc
Am A 2007;24(8):2482–4.
[238] Woods SC, Greenaway AH. Wave-front sensing by use of a Green’s function solution
to the intensity transport equation. J Opt Soc Am A 2003;20(3):508–12.
[239] Han IW. New method for estimating wavefront from curvature signal by curve
fitting. Opt Eng 1995;34(4):1232–7.
[240] Ishizuka A, Mitsuishi K, Ishizuka K. Direct observation of curvature of the wave
surface in transmission electron microscope using transport intensity equation. Ul-
tramicroscopy 2018a;194:7–14.
[241] Ishizuka A, Ishizuka K, Mitsuishi K. Boundary-artifact-free observation of mag-
netic materials using the transport of intensity equation. Microsc Microanal
2018b;24(S1):924–5.
[242] Mehrabkhani S, Wefelnberg L, Schneider T. Fourier-based solving approach
for the transport-of-intensity equation with reduced restrictions. Opt Express
2018;26(9):11458.
[243] Zhang J., Chen Q., Sun J., Tian L., Zuo C.. On a universal solution to the transport-
of-intensity equation. 2019. 1912.07371.
[244] Schmalz JA, Gureyev TE, Paganin DM, Pavlov KM. Phase retrieval using radia-
tion and matter-wave fields: Validity of teague’s method for solution of the trans-
port-of-intensity equation. Phys Rev A 2011;84(2).
[245] Shanker A, Sczyrba M, Connolly B, Kalk F, Neureuther A, Waller L. Critical assess-
ment of the transport of intensity equation as a phase recovery technique in optical
lithography. Optical Microlithography XXVII, vol 9052. International Society for
Optics and Photonics; 2014. 90521D.
[246] Sheppard CJR, Mao XQ. Three-dimensional imaging in a microscope. J Opt Soc Am
A 1989;6(9):1260–9.
[247] Hopkins HH, Barham PM. The influence of the condenser on microscopic resolu-
tion. Proc Phys Soc B 1950;63(10):737.
[248] Goodman JW. Statistical optics. wiley classics library ed. New York: Wiley; 2000.
ISBN 978-0-471-39916-2 978-0-471-01502-4
[249] Sheppard CJR. Defocused transfer function for a partially coherent microscope and
application to phase retrieval. J Opt Soc Am A 2004;21(5):828–31.
[250] McCutchen CW. Generalized Aperture and the Three-Dimensional Diffraction Im-
age. J Opt Soc Am 1964;54(2):240–4.
[251] Wilson T. Theory and Practice of Scanning Optical Microscopy. 1 edition. London
; Orlando: Academic Pr; 1984. ISBN 978-0-12-757760-9.
[252] Gu M. Advanced optical imaging theory, vol 75. Springer Science & Business Media;
2000.
[253] Pogany A, Gao D, Wilkins SW. Contrast and resolution in imaging with a microfocus
x-ray source. Rev Sci Instrum 1997;68(7):2774–82.
[254] Cloetens P, Ludwig W, Boller E, Helfen L, Salvo L, Mache R, Schlenker M. Quanti-
tative phase contrast tomography using coherent synchrotron radiation. In: Devel-
opments in X-ray tomography III, vol. 4503. International Society for Optics and
Photonics; 2002. p. 82–91.
[255] Mayo SC, Miller PR, Wilkins SW, Davis TJ, Gao D, Gureyev TE, Paganin D, Parry DJ,
Pogany A, Stevenson AW. Quantitative x-ray projection microscopy: phase-contrast
and multi-spectral imaging. J Microsc 2002;207(2):79–96.
[256] Gureyev TE, Pogany A, Paganin DM, Wilkins SW. Linear algorithms for phase re-
trieval in the Fresnel region. Opt Commun 2004a;231(1-6):53–70.
[257] Gureyev TE, Mayo S, Wilkins SW, Paganin D, Stevenson AW. Quantita-
tive in-line phase-contrast imaging with multienergy x rays. Phys Rev Lett
2001;86(25):5827–30.
[258] Wu X, Liu H. A general theoretical formalism for x-ray phase contrast imaging. J
X-ray Sci Technol 2003;11(1):33–42.
[259] Guigay JP, Langer M, Boistel R, Cloetens P. Mixed transfer function and trans-
port of intensity approach for phase retrieval in the fresnel region. Opt Lett
2007;32(12):1617–19.
[260] Langer M, Cloetens P, Guigay J-P, Peyrin F. Quantitative comparison of direct phase
retrieval algorithms in in-line phase tomography. Med Phys 2008;35(10):4556–66.
[261] Cloetens P, Barrett R, Baruchel J, Guigay J-P, Schlenker M. Phase objects in syn-
chrotron radiation hard x-ray imaging. J Phys D: Appl Phys 1996;29(1):133.
[262] Zabler S, Cloetens P, Guigay J-P, Baruchel J, Schlenker M. Optimization of phase
contrast imaging using hard x rays. Rev Sci Instrum 2005;76(7):073705.
[263] Roddier F. Curvature sensing: a diffraction theory. NOAO RD Note 1987;87-3.
[264] Paganin D, Barty A, McMahon PJ, Nugent KA. Quantitative phase-amplitude mi-
croscopy. III. the effects of noise. J Microsc 2004a;214(1):51–61.
[265] Savitzky A, Golay MJE. Smoothing and Differentiation of Data by Simplified Least
Squares Procedures. Anal Chem 1964;36(8):1627–39.
[266] Gorry PA. General least-squares smoothing and differentiation of nonuniformly
spaced data by the convolution method. Anal Chem 1991;63(5):534–6.
[267] Orfanidis SJ. Introduction to signal processing. Prentice-Hall, Inc; 1995.
Optics and Lasers in Engineering 135 (2020) 106187
[268] Luo J, Ying K, He P, Bai J. Properties of savitzky-golay digital differentiators. Digit
Signal Process 2005;15(2):122–36.
[269] Gureyev TE. Composite techniques for phase retrieval in the fresnel region. Opt
Commun 2003;220(1):49–58.
[270] Donnadieu P, Verdier M, Berthom G, Mur P. Imaging a dense nanodot assembly by
phase retrieval from TEM images. Ultramicroscopy 2004;100(1-2):79–90.
[271] Martinez-Carranza J, Falaggis K, Kozacki T. Optimum plane selection for trans-
port-of-intensity-equation-based solvers. Appl Optics 2014b;53(30):7050–8.
[272] Soto M, Acosta E. Improved phase imaging from intensity measurements in multiple
planes. Appl Optics 2007;46(33):7978–81.
[273] Jingshan Z, Claus RA, Dauwels J, Tian L, Waller L. Transport of intensity phase
imaging by intensity spectrum fitting of exponentially spaced defocus planes. Opt
Express 2014;22(9):10661–74.
[274] Martinez-Carranza J, Falaggis K, Kozacki T. Multi-filter transport of intensity equa-
tion solver with equalized noise sensitivity. Opt Express 2015;23(18):23092–107.
[275] Bostan E, Froustey E, Nilchian M, Sage D, Unser M. Variational phase imaging using
the transport-of-intensity equation. IEEE Trans Image Process 2016;25(2):807–17.
[276] Langer M, Cloetens P, Peyrin F. Regularization of phase retrieval with phase-at-
tenuation duality prior for 3-d holotomography. IEEE Trans Image Process
2010;19(9):2428–36.
[277] Zuo C, Chen Q, Qu W, Asundi A. Direct continuous phase demodulation in dig-
ital holography with use of the transport-of-intensity equation. Opt Commun
2013d;309:221–6.
[278] Tian L, Petruccelli JC, Barbastathis G. Nonlinear diffusion regularization for trans-
port of intensity phase imaging. Opt Lett 2012a;37(19):4131–3.
[279] Lee P-C. Phase retrieval method for in-line phase contrast x-ray imaging and de-
noising by regularization. Opt Express 2015;23(8):10668.
[280] Wu X, Liu H. X-ray cone-beam phase tomography formulas based on phase-atten-
uation duality. Opt Express 2005;13(16):6000–14.
[281] Wolf E, et al. Introduction to the Theory of Coherence and Polarization of Light.
Cambridge University Press; 2007.
[282] Mandel L, Wolf E. Optical coherence and quantum optics. Cambridge university
press; 1995.
[283] Zernike F. The concept of degree of coherence and its application to optical prob-
lems. Physica 1938;5(8):785–95.
[284] Mandel L, Wolf E. Spectral coherence and the concept of cross-spectral purity∗ . J
Opt Soc Am 1976;66(6):529–35.
[285] Chatfield C. The analysis of time series: theory and practice. Springer; 2013.
[286] Wolf E. New theory of partial coherence in the space-frequency domain. Part I:
spectra and cross spectra of steady-state sources. J Opt Soc Am 1982;72(3):343–51.
[287] van Cittert PH. Die wahrscheinliche schwingungsverteilung in einer von
einer lichtquelle direkt oder mittels einer linse beleuchteten ebene. Physica
1934;1(1):201–10.
[288] Zernike F. Diffraction and optical image formation. Proc Phys Soc 1948;61(2):158.
[289] Gori F. Directionality and spatial coherence. Opt Acta Int J Opt
1980;27(8):1025–34.
[290] Wolf E. New spectral representation of random sources and of the partially coherent
fields that they generate. Opt Commun 1981;38(1):3–6.
[291] Starikov A, Wolf E. Coherent-mode representation of Gaussian Schell-model
sources and of their radiation fields. J Opt Soc Am 1982;72(7):923–8.
[292] Gori F, Santarsiero M, Borghi R, Guattari G. Intensity-based modal analysis of par-
tially coherent beams with Hermite-Gaussian modes. Opt Lett 1998;23(13):989–91.
[293] Allen L, Beijersbergen MW, Spreeuw RJC, Woerdman JP. Orbital angular momen-
tum of light and the transformation of Laguerre-Gaussian laser modes. Phys Rev A
1992;45(11):8185–9.
[294] Schell A. A technique for the determination of the radiation pattern of a partially
coherent aperture. IEEE Trans Antennas Propag 1967;15(1):187–8.
[295] Nugent KA. A generalization of Schell’s theorem. Opt Commun 1990;79(5):267–9.
[296] Nugent KA. Coherent methods in the x-ray sciences. Adv Phys 2010;59(1):1–99.
[297] Siegman AE. New developments in laser resonators. In: Optical resonators, vol.
1224. International Society for Optics and Photonics; 1990. p. 2–14.
[298] Wigner EP. On the quantum correction for thermodynamic equilibrium. In: Part I:
Physical Chemistry. Part II: Solid State Physics. Springer; 1997. p. 110–20.
[299] Dolin L. Beam description of weakly-inhomogeneous wave fields. Izv Vyssh Uchebn
Zaved Radiofiz 1964;7:559–63.
[300] Walther A. Radiometry and coherence. J Opt Soc Am 1968;58(9):1256–9.
[301] Walther A. Radiometry and coherence. J Opt Soc Am 1973;58(9):1256–9.
[302] Bastiaans MJ. The wigner distribution function applied to optical signals and sys-
tems. Opt Commun 1978;25(1):26–30.
[303] Bastiaans MJ. Application of the wigner distribution function to partially coherent
light. J Opt Soc Am A 1986;3(8):1227–38.
[304] Bastiaans MJ. Wigner distribution function and its application to first-order optics.
J Opt Soc Am 1979a;69(12):1710–16.
[305] Woodward PM. Probability and information theory, with applications to radar.
Pergamon science series. Electronics and waves, a series of monographs. New York:
McGraw-Hill; 1953.
[306] Papoulis A. Ambiguity function in fourier optics. J Opt Soc Am 1974;64(6):779–88.
[307] Testorf M, Hennelly B, Ojeda-Castaneda J. Phase-Space Optics: Fundamentals and
Applications. 1 edition. New York: McGraw-Hill Education; 2009.
[308] Boashash B. Estimating and interpreting the instantaneous frequency of a signal. i.
fundamentals. Proc IEEE 1992;80(4):520–38.
[309] McBRIDE AC, Kerr FH. On namias’s fractional fourier transforms. IMA J Appl Math
1987;39(2):159–75.
[310] Lohmann AW. Image rotation, wigner rotation, and the fractional fourier transform.
J Opt Soc Am A 1993;10(10):2181–6.
C. Zuo, J. Li and J. Sun et al.
[311] Mendlovic D, Ozaktas HM. Fractional Fourier transforms and their optical imple-
mentation: I. J Opt Soc Am A 1993;10(9):1875–81.
[312] Ozaktas HM, Kutay MA. The fractional fourier transform. In: 2001 European Con-
trol Conferenc (ECC); 2001. p. 1477–83.
[313] Lohmann AW, Soffer BH. Relationships between the Radon-Wigner and fractional
Fourier transforms. J Opt Soc Am A 1994;11(6):1798–801.
[314] Raymer MG, Beck M, McAlister D. Complex wave-field reconstruction using
phase-space tomography. Phys Rev Lett 1994;72(8):1137–40.
[315] McAlister DF, Beck M, Clarke L, Mayer A, Raymer MG. Optical phase retrieval
by phase-space tomography and fractional-order Fourier transforms. Opt Lett
1995;20(10):1181–3.
[316] McKee CB, O’Shea PG, Madey JMJ. Phase space tomography of relativistic electron
beams. Nuclear Instruments and Methods in Physics Research Section A: Accelera-
tors, Spectrometers, Detectors and Associated Equipment 1995;358(1):264–7.
[317] Tian L, Lee J, Oh SB, Barbastathis G. Experimental compressive phase space to-
mography. Opt Express 2012b;20(8):8296–308.
[318] Nazarathy M, Shamir J. First-order optics-a canonical operator representation: loss-
less systems. J Opt Soc Am 1982;72(3):356–64.
[319] Gerrard A, Burch JM. Introduction to Matrix Methods in Optics. Courier Corpora-
tion; 1975. ISBN 978-0-486-68044-6.
[320] Collins SA. Lens-system diffraction integral written in terms of matrix optics. J Opt
Soc Am 1970;60(9):1168–77.
[321] Brenner K-H, Lohmann AW. Wigner distribution function display of complex 1d
signals. Opt Commun 1982;42(5):310–14.
[322] Bastiaans MJ. Transport equations for the wigner distribution function in an inho-
mogeneous and dispersive medium. Opt Acta Int J Opt 1979b;26(11):1333–44.
[323] Winston R, Welford WT. Geometrical vector flux and some new nonimaging con-
centrators. J Opt Soc Am 1979;69(4):532–6.
[324] Naik DN, Pedrini G, Takeda M, Osten W. Spectrally resolved incoherent hologra-
phy: 3d spatial and spectral imaging using a mach-zehnder radial-shearing inter-
ferometer. Opt Lett 2014;39(7):1857.
[325] Naik DN, Pedrini G, Osten W. Recording of incoherent-object hologram as com-
plex spatial coherence function using Sagnac radial shearing interferometer and a
Pockels cell. Opt Express 2013;21(4):3990–5.
[326] Iaconis C, Walmsley IA. Direct measurement of the two-point field correlation func-
tion. Opt Lett 1996;21(21):1783–5.
[327] Marks DL, Stack RA, Brady DJ. Three-dimensional coherence imaging in the fresnel
domain. Appl Optics 1999;38(8):1332–42.
[328] Kak AC, Slaney M, Wang G. Principles of computerized tomographic imaging. Med
Phys 2002;29(1). 107–107
[329] Radon J. On the determination of functions from their integral values along certain
manifolds. IEEE T Med Imaging 1986;5(4):170–6.
[330] Tu J, Tamura S. Wave field determination using tomography of the ambiguity func-
tion. Phys Rev E 1997;55(2):1946–9.
[331] Dragoman D, Dragoman M, Brenner KH. Tomographic amplitude and phase recov-
ery of vertical-cavity surface-emitting lasers by use of the ambiguity function. Opt
Lett 2002;27(17):1519–21.
[332] Liu X, Brenner KH. Reconstruction of two-dimensional complex amplitudes from
intensity measurements. Opt Commun 2003;225(1-3):19–30.
[333] Testorf ME, Semichaevsky A. Phase retrieval and phase-space tomography from
incomplete data sets. In: Image Reconstruction from Incomplete Data III, vol. 5562.
International Society for Optics and Photonics; 2004. p. 38–49.
[334] Bartelt HO, Brenner K-H, Lohmann AW. The wigner distribution function and its
optical production. Opt Commun 1980;32(1):32–8.
[335] Zhang Z, Levoy M. Wigner distributions and how they relate to the light field. In:
2009 IEEE International Conference on Computational Photography (ICCP); 2009.
p. 1–10. doi:10.1109/ICCPHOT.2009.5559007.
[336] Waller L, Situ G, Fleischer JW. Phase-space measurement and coherence synthesis
of optical beams. Nat Photonics 2012;6(7):474–9.
[337] Ng R, Levoy M, Brdif M, Duval G, Horowitz M, Hanrahan P. Light field photography
with a hand-held plenoptic camera. Comput Sci Tech Rep 2005;2:1–11.
[338] Tian L, Zhang Z, Petruccelli JC, Barbastathis G. Wigner function measurement using
a lenslet array. Opt Express 2013;21(9):10511.
[339] Stoklasa B, Motka L, Rehacek J, Hradil Z, Snchez-Soto LL. Wavefront sensing re-
veals optical coherence. Nat Commun 2014;5:ncomms4275.
[340] Banaszek K, Wdkiewicz K. Direct probing of quantum phase space by photon count-
ing. Phys Rev Lett 1996;76(23):4344–7.
[341] Chapman HN. Phase-retrieval x-ray microscopy by wigner-distribution deconvolu-
tion. Ultramicroscopy 1996;66(3):153–72.
[342] Chakraborty T, Petruccelli JC. Source diversity for transport of intensity phase
imaging. Opt Express 2017;25(8):9122–37.
[343] Hamilton DK, Sheppard C. Differential phase contrast in scanning optical mi-
croscopy. J Microsc-oxford 1984;133(1):27–39.
[344] Tian L, Waller L. Quantitative differential phase contrast imaging in an LED array
microscope. Opt Express 2015b;23(9):11394–403.
[345] Fan Y, Sun J, Chen Q, Pan X, Tian L, Zuo C. Optimal illumination scheme
for isotropic quantitative differential phase contrast microscopy. Photon Res
2019;7(8):890–904. doi:10.1364/PRJ.7.000890.
[346] Iglesias I. Pyramid phase microscopy. Opt Lett 2011;36(18):3636–8.
[347] Parthasarathy AB, Chu KK, Ford TN, Mertz J. Quantitative phase imaging using a
partitioned detection aperture. Opt Lett 2012;37(19):4062–4.
[348] Zuo C, Sun J, Feng S, Zhang M, Chen Q. Programmable aperture microscopy: A
computational method for multi-modal phase contrast and light field imaging. Opt
Laser Eng 2016b;80:24–31.
Optics and Lasers in Engineering 135 (2020) 106187
[349] Gureyev TE, Nesterets YI, Paganin DM, Pogany A, Wilkins SW. Linear algorithms
for phase retrieval in the fresnel region. 2. partially coherent illumination. Opt
Commun 2006;259(2):569–80.
[350] Zysk AM, Schoonover RW, Carney PS, Anastasio MA. Transport of intensity and
spectrum for partially coherent fields. Opt Lett 2010;35(13):2239–41.
[351] Petruccelli JC, Tian L, Barbastathis G. The transport of intensity equation for
optical path length recovery using partially coherent illumination. Opt Express
2013;21(12):14430–41.
[352] Gureyev TE, Paganin DM, Stevenson AW, Mayo SC, Wilkins SW. Generalized
eikonal of partially coherent beams and its use in quantitative imaging. Phys Rev
Lett 2004b;93(6):068103.
[353] Ng R. Fourier slice photography. In: ACM Transactions on Graphics (TOG), vol. 24.
ACM; 2005. p. 735–44.
[354] Sheppard CJR. Partially coherent microscope imaging system in phase space: effect
of defocus and phase reconstruction. J Opt Soc Am A 2018;35(11):1846.
[355] Paganin DM, Labriet H, Brun E, Berujon S. Single-image geometric-flow x-ray
speckle tracking. Physl Rev A 2018;98(5):053813.
[356] Lu L, Sun J, Zhang J, Fan Y, Chen Q, Zuo C. Quantitative phase imaging camera
with a weak diffuser. Front Phys 2019a;7:77.
[357] Zdora M-C. State of the art of x-ray speckle-based phase-contrast and dark-field
imaging. J Imaging 2018;4(5):60.
[358] Chellappan KV, Erden E, Urey H. Laser-based displays: a review. Appl Optics
2010;49(25):F79–98.
[359] Liba O, Lew MD, SoRelle ED, Dutta R, Debasish S, Moshfeghi DM, Chu S, de La
Zerda A. Speckle-modulating optical coherence tomography in living mice and hu-
mans. Nat Commun 2017;8:15845.
[360] Bates R. Astronomical speckle imaging. Phys Rep 1982;90(4):203–97.
[361] Høgmoen K, Pedersen HM. Measurement of small vibrations using electronic
speckle pattern interferometry: theory. J Opt Soc Am 1977;67(11):1578–83.
[362] Boas DA, Dunn AK. Laser speckle contrast imaging in biomedical optics. J Biomed
Opt 2010;15(1):011109.
[363] Lu L, Sun J, Zhang J, Fan Y, Chen Q, Zuo C. Speckle quantitative phase imaging
based on coherence effect compensation. Seventh International Conference on Op-
tical and Photonic Engineering (icOPEN 2019), vol 11205. International Society
for Optics and Photonics; 2019b. 112050I.
[364] Orth A, Crozier KB. Light field moment imaging. Opt Lett 2013;38(15):2666–8.
[365] Liu J, Xu T, Yue W, Sun J, Situ G. Light-field moment microscopy with noise re-
duction. Opt Express 2015;23(22):29154–62.
[366] Jenkins MH, Long JM, Gaylord TK. Multifilter phase imaging with partially coher-
ent light. Appl Optics 2014;53(16):D29–39.
[367] Jenkins MH, Gaylord TK. Quantitative phase microscopy via optimized inversion
of the phase optical transfer function. Appl Optics 2015a;54(28):8566–79.
[368] Bao Y, Gaylord TK. Two improved defocus quantitative phase imaging methods:
discussion. J Opt Soc Am A 2019;36(12):2104–14.
[369] Martinez-Carranza J, Falaggis K, Kozacki T. Enhanced lateral resolution for phase
retrieval based on the transport of intensity equation with tilted illumination. In:
Popescu G, Park Y, editors. Quantitative Phase Imaging II, vol. 9718. SPIE; 2016.
p. 65–74. doi:10.1117/12.2212876. International Society for Optics and Photonics
[370] Zhu Y, Shanker A, Tian L, Waller L, Barbastathis G. Low-noise phase imaging by
hybrid uniform and structured illumination transport of intensity equation. Opt
Express 2014a;22(22):26696–711. doi:10.1364/OE.22.026696.
[371] Zhu Y, Zhang Z, Barbastathis G. Phase imaging for absorptive phase objects using
hybrid uniform and structured illumination transport of intensity equation. Opt
Express 2014b;22(23):28966–76. doi:10.1364/OE.22.028966.
[372] Chakraborty T, Petruccelli JC. Optical convolution for quantitative phase retrieval
using the transport of intensity equation. Appl Optics 2018;57(1):A134–41.
[373] Bao Y, Dong GC, Gaylord TK. Weighted-least-squares multi-filter phase imaging
with partially coherent light: characteristics of annular illumination. Appl Optics
2019;58:137.
[374] Kak AC, Slaney M. Principles of Computerized Tomographic Imaging. SIAM; 1988.
ISBN 978-0-89871-927-7.
[375] Wolf E. Three-dimensional structure determination of semi-transparent objects
from holographic data. Opt Commun 1969;1(4):153–6.
[376] Haeberl O, Belkebir K, Giovaninni H, Sentenac A. Tomographic diffractive mi-
croscopy: basics, techniques and perspectives. J Mod Opt 2010;57(9):686–99.
[377] Rappaz B, Marquet P, Cuche E, Emery Y, Depeursinge C, Magistretti PJ. Measure-
ment of the integral refractive index and dynamic cell morphometry of living cells
with digital holographic microscopy. Opt Express 2005;13(23):9361–73.
[378] Lauer V. New approach to optical diffraction tomography yielding a vector equa-
tion of diffraction tomography and a novel tomographic microscope. J Microsc
2002;205(Pt 2):165–76.
[379] Choi W. Tomographic phase microscopy and its biological applications. 3D Res
2012;3(4).
[380] Jenkins MH, Gaylord TK. Three-dimensional quantitative phase imaging via tomo-
graphic deconvolution phase microscopy. Appl Optics 2015b;54(31):9213–27.
[381] Soto JM, Rodrigo JA, Alieva T. Label-free quantitative 3d tomographic imaging for
partially coherent light microscopy. Opt Express 2017;25(14):15699–712.
[382] Devaney AJ. Inverse-scattering theory within the Rytov approximation. Opt Lett
1981;6(8):374–6.
[383] Devaney AJ. A filtered backpropagation algorithm for diffraction tomography. Ul-
trason Imaging 1982;4(4):336–50.
[384] Sung Y, Choi W, Fang-Yen C, Badizadegan K, Dasari RR, Feld MS. Optical diffraction
tomography for high resolution live cell imaging. Opt Express 2009;17(1):266–77.
[385] Fiddy MA. Inversion of optical scattered field data. J Phys D: Appl Phys
1986;19(3):301.
C. Zuo, J. Li and J. Sun et al.
[386] Chen B, Stamnes JJ. Validity of diffraction tomography based on the first born and
the first rytov approximations. Appl Optics 1998;37(14):2996–3006.
[387] Dändliker R, Weiss K. Reconstruction of the three-dimensional refractive index
from scattered waves. Opt Commun 1970;1(7):323–8.
[388] Bao Y, Gaylord TK. Quantitative phase imaging method based on an analyti-
cal nonparaxial partially coherent phase optical transfer function. J Opt Soc Am
2016;33(11):2125.
[389] Bao Y, Gaylord TK. Clarification and unification of the obliquity factor in diffraction
and scattering theories: discussion. J Opt Soc Am A 2017;34:1738–45.
[390] Sheppard CJR, Matthews HJ. Imaging in high-aperture optical systems. J Opt Soc
Am A 1987;4(8):1354–60.
[391] Sheppard CJR, Kawata Y, Kawata S, Gu M. Three-dimensional transfer functions
for high-aperture systems. J Opt Soc Am A 1994;11(2):593–8.
[392] Choi W, Fang-Yen C, Badizadegan K, Oh S, Lue N, Dasari RR, Feld MS. Tomographic
phase microscopy. Nat Methods 2007;4(9):717.
[393] Kim K, Yoon H, Diez-Silva M, Dao M, Dasari RR, Park Y. High-resolution three-
-dimensional imaging of red blood cells parasitized by plasmodium falciparum
and in situ hemozoin crystals using optical diffraction tomography. J Biomed Opt
2013;19(1):011005.
[394] Cotte Y, Toy F, Jourdain P, Pavillon N, Boss D, Magistretti P, Marquet P, De-
peursinge C. Marker-free phase nanoscopy. Nat Photonics 2013;7(2):113–17.
[395] Charrière F, Marian A, Montfort F, Kuehn J, Colomb T, Cuche E, Marquet P, De-
peursinge C. Cell refractive index tomography by digital holographic microscopy.
Opt Lett 2006;31(2):178.
[396] Charrire F, Pavillon N, Colomb T, Depeursinge C, Heger TJ, Mitchell EAD, Mar-
quet P, Rappaz B. Living specimen tomography by digital holographic microscopy:
morphometry of testate amoeba. Opt Express 2006;14(16):7005.
[397] Soto JM, Rodrigo JA, Alieva T. Partially coherent illumination engineering for en-
hanced refractive index tomography. Opt Lett 2018;43(19):4699.
[398] Blanchard PM, Fisher DJ, Woods SC, Greenaway AH. Phase-diversity wave-front
sensing with a distorted diffraction grating. Appl Optics 2000;39(35):6649–55.
[399] Di Martino JM, Ayubi GA, Dalchiele EA, Alonso JR, Fernández A, Flores JL, Per-
ciante CD, Ferrari JA. Single-shot phase recovery using two laterally separated
defocused images. Opt Commun 2013;293:1–3.
[400] Lee JW, Ku J, Waller L, Barbastathis G. Transport of intensity imaging applied to
quantitative optical phase tomography. In: Digital Holography and Three-Dimen-
sional Imaging. Optical Society of America; 2011. p. DTuD2.
[401] Nguyen T, Nehmetallah G, Tran D, Darudi A, Soltani P. Fully automated,
high speed, tomographic phase object reconstruction using the transport of
intensity equation in transmission and reflection configurations. Appl Optics
2015;54(35):10443–53.
[402] Nguyen T, Nehmetallah G. Non-interferometric tomography of phase objects using
spatial light modulators. J Imaging 2016;2(4):30.
[403] Phillips KG, Velasco CR, Li J, Kolatkar A, Luttgen M, Bethel K, Duggan B, Kuhn P,
McCarty OJT. Optical quantification of cellular mass, volume, and density of circu-
lating tumor cells identified in an ovarian cancer patient. Front Oncol 2012a;2:72.
[404] Phillips KG, Jacques SL, McCarty OJT. Measurement of single cell refractive index,
dry mass, volume, and density using a transillumination microscope. Phys Rev Lett
2012b;109(11):118105.
[405] Paganin D, Gureyev TE, Mayo SC, Stevenson AW, Nesterets YI, Wilkins SW. X-ray
omni microscopy. J Microsc 2004b;214(3):315–27.
[406] Li Y, Di J, Ma C, Zhang J, Zhong J, Wang K, Xi T, Zhao J. Quantitative phase
microscopy for cellular dynamics based on transport of intensity equation. Opt
Express 2018c;26(1):586.
[407] Zheng J, Zuo C, Gao P, Nienhaus GU. Dual-mode phase and fluorescence imaging
with a confocal laser scanning microscope. Opt Lett 2018;43(22):5689.
[408] Zheng J-J, Yao B-L, Shao X-P. Nonresonant background suppression in wide-field
coherent anti-stokes raman scattering microscopy with transport of intensity equa-
tion based phase imaging. Acta Phys Sin 2017;66(11):114206.
[409] Zheng J, Akimov D, Heuke S, Schmitt M, Yao B, Ye T, Lei M, Gao P, Popp J. Vibra-
tional phase imaging in wide-field CARS for nonresonant background suppression.
Opt Express 2015;23(8):10756–63.
[410] Liao J, Wang Z, Zhang Z, Bian Z, Guo K, Nambiar A, Jiang Y, Jiang S, Zhong J,
Choma M, Zheng G. Dual light-emitting diode-based multichannel microscopy
for whole-slide multiplane, multispectral and phase imaging. J Biophotonics
2018;11(2):e201700075.
[411] Li Y, Di J, Wu W, Shang P, Zhao J. Quantitative investigation on morphology and in-
tracellular transport dynamics of migrating cells. Appl Optics 2019a;58(34):G162.
[412] Ma L, Rajshekhar G, Wang R, Bhaduri B, Sridharan S, Mir M, Chakraborty A, Iyer R,
Prasanth S, Millet L, Gillette MU, Popescu G. Phase correlation imaging of unla-
beled cell dynamics. Sci Rep 2016;6(1):32702.
[413] Bian Y, Yao Y, Liu Q, Xu Y, Kuang C, Li H, Liu X. Assessment of tissues- inhomoge-
neous optical properties based on a portable microscope under partially coherent
illumination. Opt Commun 2019;434:145–51.
[414] Rajput SK, Kumar M, Quan X, Morita M, Furuyashiki T, Awatsuji Y, Tajahuerce E,
Matoba O. Three-dimensional fluorescence imaging using the transport of intensity
equation. J Biomed Opt 2019;25(3):032004.
[415] Mandula O, Kleman J-P, Lacroix F, Allier C, Fiole D, Herv L, Blandin P, Kraemer DC,
Morales S. Phase and fluorescence imaging with a surprisingly simple microscope
based on chromatic aberration. Opt Express 2020;28(2):2079–90.
[416] Wittkopp JM, Khoo TC, Carney S, Pisila K, Bahreini SJ, Tubbesing K, Tubbesing K,
Mahajan S, Sharikova A, Petruccelli JC, Khmaladze A. Comparative phase imaging
of live cells by digital holographic microscopy and transport of intensity equation
methods. Opt Express 2020;28(5):6123–33.
Optics and Lasers in Engineering 135 (2020) 106187
[417] Rodrigo JA, Soto JM, Alieva T. Fast label-free microscopy technique for 3d dynamic
quantitative imaging of living cells. Biomed Opt Express 2017;8(12):5507–17.
[418] Ma X, Zhou B, Su Z, Zhang Z, Peng J, Zhong J. Label-free 3d imaging of weakly
absorbing samples using spatially-incoherent annular illumination microscopy. Ul-
tramicroscopy 2019;200:97–104.
[419] Li J, Matlock A, Li Y, Chen Q, Zuo C, Tian L. High-speed in vitro intensity diffraction
tomography. Adv Photon 2019b;1(6):066004.
[420] Soto JM, Mas A, Rodrigo JA, Alieva T, Domnguez-Bernal G. Label-free bioanalysis
of leishmania infantum using refractive index tomography with partially coherent
illumination. J Biophotonics 2019;12(9):e201900030.
[421] Frank J, Matrisch J, Horstmann J, Altmeyer S, Wernicke G. Refractive index de-
termination of transparent samples by noniterative phase retrieval. Appl Optics
2011;50(4):427–33.
[422] Shomali R, Darudi A, Nasiri S. Application of irradiance transport equation in as-
pheric surface testing. Opt - Int J Light Electron Opt 2012;123(14):1282–6.
[423] Pan S, Ma J, Zhu R, Ba T, Zuo C, Chen F, Dou J, Wei C, Zhou W. Real-time complex
amplitude reconstruction method for beam quality m 2 factor measurement. Opt
Express 2017;25(17):20142–55.
[424] Bunsen M, Tateyama S. Detection method for the complex amplitude of a signal
beam with intensity and phase modulation using the transport of intensity equation
for holographic data storage. Opt Express 2019;27(17):24029–42.
[425] Seo S, Su T-W, Derek KT, Erlinger A, Ozcan A. Lensfree holographic imaging for
on-chip cytometry and diagnostics. Lab Chip 2009;9(6):777–87.
[426] Greenbaum A, Ozcan A. Maskless imaging of dense samples using pixel su-
per-resolution based multi-height lensfree on-chip microscopy. Opt Express
2012;20(3):3129.
[427] Luo W, Zhang Y, Grcs Z, Feizi A, Ozcan A. Propagation phasor approach for holo-
graphic image reconstruction. Sci Rep 2016;6:srep22738.
[428] Bishara W, Su T-W, Coskun AF, Ozcan A. Lensfree on-chip microscopy over a wide
field-of-view using pixel super-resolution. Opt Express 2010;18(11):11181.
[429] Greenbaum A, Zhang Y, Feizi A, Chung P-L, Luo W, Kandukuri SR, Ozcan A. Wide–
field computational imaging of pathology slides using lens-free on-chip microscopy.
Sci Transl Med 2014;6(267). 267ra175–267ra175
[430] Bishara W, Sikora U, Mudanyali O, Su T-W, Yaglidere O, Luckhart S, Ozcan A.
Holographic pixel super-resolution in portable lensless on-chip microscopy using a
fiber-optic array. Lab Chip 2011;11(7):1276–9.
[431] Luo W, Zhang Y, Feizi A, Grcs Z, Ozcan A. Pixel super-resolution using wavelength
scanning. Light Sci Appl 2015;5(4):e16060.
[432] Zhang J, Chen Q, Li J, Sun J, Zuo C. Lensfree dynamic super-resolved phase imaging
based on active micro-scanning. Opt Lett 2018;43(15):3714.
[433] Zhang Y, Wu Y, Zhang Y, Ozcan A. Color calibration and fusion of lens-free and mo-
bile-phone microscopy images for high-resolution and accurate color reproduction.
Sci Rep 2016;6(1).
[434] Zhang J, Sun J, Chen Q, Zuo C. Resolution analysis in a lens-free on-chip digital
holographic microscope. IEEE T Comput Imag 2020;6:697–710.
[435] Wang H, Chen N, Zheng S, Liu J, Situ G. Fast and high-resolution light field acqui-
sition using defocus modulation. Appl Optics 2018;57(1):A250–6.
[436] Chen N, Zuo C, Lam EY, Lee B. 3d imaging based on depth measurement technolo-
gies. Sensors 2018;18(11):3711.
[437] Pandey N, Ghosh A, Khare K. Two-dimensional phase unwrapping using the trans-
port of intensity equation. Appl Optics 2016;55(9):2418.
[438] Zuo C. Connections between transport of intensity equation and two-dimensional
phase unwrapping. arXiv preprint arXiv:170403950 2017.
[439] Martinez-Carranza J, Falaggis K, Kozacki T. Fast and accurate phase-unwrap-
ping algorithm based on the transport of intensity equation. Appl Optics
2017;56(25):7079–88.
[440] Zhou W-J, Guan X, Liu F, Yu Y, Zhang H, Poon T-C, Banerjee PP. Phase
retrieval based on transport of intensity and digital holography. Appl Optics
2018;57(1):A229–34.
[441] Cheng H, Wang J, Gao Y, Zhang Q, Wei S. Phase unwrapping based on trans-
port-of-intensity equation with two wavelengths. Opt Eng 2019;58(5):1–7.
[442] Zhao Z, Zhang H, Xiao Z, Du H, Zhuang Y, Fan C, Zhao H. Robust 2d phase un-
wrapping algorithm based on the transport of intensity equation. Meas Sci Technol
2019;30(1):015201.
[443] Zhang C, He W, Wu J, Peng X. Optical cryptosystem based on phase-trun-
cated fresnel diffraction and transport of intensity equation. Opt Express
2015;23(7):8845–54.
[444] Yoneda N, Saita Y, Komuro K, Nobukawa T, Nomura T. Transport-of-intensity
holographic data storage based on a computer-generated hologram. Appl Optics
2018;57(30):8836–40.
[445] Rajput SK, Matoba O. Security-enhanced optical voice encryption in various do-
mains and comparative analysis. Appl Optics 2019;58(11):3013–22.
[446] Sui L, Zhao X, Huang C, Tian A, Anand A. An optical multiple-image authentication
based on transport of intensity equation. Opt Lasers Eng 2019;116:116–24.
[447] Yang Z, Zhan Q. Single-shot smartphone-based quantitative phase imaging using a
distorted grating. PloS one 2016;11(7):e0159596.
[448] Meng X, Huang H, Yan K, Tian X, Yu W, Cui H, Kong Y, Xue L, Liu C, Wang S.
Smartphone based hand-held quantitative phase microscope using the transport of
intensity equation method. Lab Chip 2017;17(1):104–9.
[449] Zuo C, Chen Q, Asundi A. Comparison of digital holography and transport of in-
tensity for quantitative phase contrast imaging. In: Fringe 2013. Springer; 2014e.
p. 137–42.
[450] Lubk A, Guzzinati G, Börrnert F, Verbeeck J. Transport of intensity phase retrieval
of arbitrary wave fields including vortices. Phys Rev Lett 2013;111(17):173902.
[451] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521(7553):436–44.
C. Zuo, J. Li and J. Sun et al.
[452] Sinha A, Lee J, Li S, Barbastathis G. Lensless computational imaging through deep
learning. Optica 2017;4(9):1117.
[453] Rivenson Y, Zhang Y, Gnaydn H, Teng D, Ozcan A. Phase recovery and holo-
graphic image reconstruction using deep learning in neural networks. Light-Sci
Appl 2018;7(2):17141.
Optics and Lasers in Engineering 135 (2020) 106187
[454] Feng S, Chen Q, Gu G, Tao T, Zhang L, Hu Y, Yin W, Zuo C. Fringe pattern analysis
using deep learning. Adv Photon 2019;1(2):025001.
[455] Barbastathis G, Ozcan A, Situ G. On the use of deep learning for computational
imaging. Optica 2019;6(8):921.